An ultrasonic image offline acquisition and dynamic synchronous processing method and system

Abstract

The application discloses an ultrasonic image offline collection and dynamic synchronous processing method and system, which realizes offline collection and dynamic synchronous processing of ultrasonic images under network interruption conditions through six function modules of network state monitoring, online data synchronization, image collection and processing, offline mode switching, offline diagnosis processing and incremental synchronization. The method comprises the following steps: acquiring network connection state data and determining a working mode; pulling task and template data and caching in an online state; acquiring multi-modal image data and double-writing storage; detecting network interruption and switching to an offline mode; performing AI analysis on the offline collected images and generating a report; detecting network recovery and performing incremental data synchronization. The application solves the business interruption problem caused by the dependence of the traditional PACS system on the network, improves the network adaptability and robustness of the system, reduces the operation and maintenance cost, and shortens the system fault recovery time.

Description

Technical Field

[0001] This invention relates to the field of medical information system technology, specifically to a method and system for offline acquisition and dynamic synchronous processing of ultrasound images. Background Technology

[0002] With the rapid development of medical informatization, medical image acquisition and processing systems have become an indispensable part of modern hospitals. However, existing medical imaging systems suffer from the following technical problems:

[0003] First, existing PACS systems require ultrasound equipment, workstations, and servers to be connected to the network in real time (e.g., image acquisition and transmission modules require DICOM online services). If the network fluctuates or the server crashes, it will lead to the inability to register patient information; the inability to acquire / store images; and the inability to complete examinations and write reports. This causes a complete standstill in departmental operations, resulting in wasted medical resources, patient backlogs, and an inability to respond to emergency ultrasound calls.

[0004] Secondly, traditional solutions only support basic data caching (such as temporary storage of patient information), but cannot complete image acquisition offline, cannot call structured templates to generate reports, and cannot print reports; they can only perform fragmented operations and cannot form a closed loop of "acquisition → processing → reporting". After the network is restored, data needs to be manually synchronized, which can easily lead to report loss or version conflicts.

[0005] Finally, existing offline solutions (such as standalone ultrasound workstations) suffer from irreversible data isolation. Images and reports are stored only on the local client, requiring manual export / import of data after network recovery. Data for the same patient is scattered across different workstations, leading to incomplete diagnostic evidence (such as inability to match historical examinations) and violating HIPAA regulations regarding data traceability.

[0006] Therefore, there is an urgent need for a method and system for offline acquisition and dynamic synchronization processing of ultrasound images that can maintain business continuity during network interruptions and automatically synchronize data after network recovery. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for offline acquisition and dynamic synchronous processing of ultrasound images, so as to solve the technical problems in the prior art, such as service interruption caused by network dependence, data fragmentation caused by lack of offline function, and irreversible data isolation in offline solutions.

[0008] To achieve the above objectives, the present invention provides a method for offline acquisition and dynamic synchronous processing of ultrasound images, comprising the following steps:

[0009] Obtain network connection status data between the client and the server, continuously monitor and perform connectivity detection on the network connection status data, determine the current network working mode, and obtain the network working mode result;

[0010] When the network working mode result is online, the pending inspection task data and report template data are retrieved from the server, and the pending inspection task data and the report template data are stored in the local cache database to establish an online data synchronization mechanism.

[0011] Based on the online data synchronization mechanism, multimodal image data acquired by ultrasound equipment is obtained. The multimodal image data is parsed using the DICOM protocol and converted into a format to obtain converted image data. The converted image data is then written and stored in both a local temporary library and the server PACS system to generate local image backup data.

[0012] Based on the changes in the network connection status data, a network interruption event is detected and an offline mode switch is triggered. The local SQLite database is invoked to take over the data storage service, the local image backup data is activated, and the local image acquisition and processing process is started.

[0013] The ultrasound image data acquired in offline mode is analyzed and lesion is identified using local AI algorithms. The offline diagnostic report is generated by combining the report template data in the local cache database, thus completing the offline business loop.

[0014] The network connection status data is continuously monitored to detect network recovery events and trigger the data synchronization process. The data to be synchronized during offline periods is identified through a difference comparison algorithm, and incremental data is uploaded to the server to complete the offline acquisition and dynamic synchronization processing of multimodal fusion ultrasound images.

[0015] Preferably, based on the online data synchronization mechanism, multimodal image data acquired by the ultrasound equipment is obtained; the multimodal image data is parsed using the DICOM protocol and converted to obtain converted image data; and the converted image data is dual-written and stored in a local temporary library and a server PACS system to generate local image backup data, including:

[0016] A three-dimensional sparse tensor structure is constructed for the multimodal image data, where the first dimension represents the spatial location coordinates, the second dimension represents the time series, and the third dimension represents the modal features. Regions of interest are extracted and sparse tensor representations are generated.

[0017] Based on the sparse tensor representation, the local search principal component analysis algorithm is applied to perform feature dimensionality reduction. The optimal principal component direction is found through gradient projection and soft thresholding to generate compressed feature data.

[0018] The compressed feature data is quality-assessed, and the variance explanation rate and the retention rate of key diagnostic information are calculated. When the variance explanation rate reaches a preset validity threshold, the compression validity is confirmed, and the compressed image data is stored in the local temporary library.

[0019] Based on the compressed feature data, image statistical features, texture descriptors, and morphological parameters are calculated to generate a preprocessed feature library required for offline AI analysis, thus obtaining the local image backup data.

[0020] Preferably, the step of detecting network interruption events and triggering offline mode switching based on changes in the network connection status data, calling the local SQLite database to take over data storage services, activating the local image backup data, and starting the local image acquisition and processing flow includes:

[0021] Each frame in the dynamic ultrasound image sequence acquired in offline mode is constructed as a node in a graph structure, and the spatiotemporal correlation between frames is calculated as the edge weight to establish a temporal correlation graph.

[0022] Connectivity component analysis is performed on the temporal correlation graph to identify missing frames, damaged frames, and temporal anomaly regions in the image sequence, and the number and topological characteristics of connected components are counted to obtain the connectivity component analysis results.

[0023] Based on the connected component analysis results and combined with the preprocessed feature library in the local image backup data, the optimal graph reconstruction algorithm is applied to repair the integrity of the image sequence. Missing data segments are filled in by feature interpolation and periodic reconstruction methods to obtain the reconstructed image sequence.

[0024] The reconstructed image sequence is subjected to multi-level consistency verification, including pixel-level grayscale continuity verification, feature-level anatomical structure continuity verification, and semantic-level physiological parameter rationality verification, to generate complete localized image data.

[0025] The complete localized image data is stored hierarchically according to patient ID, examination items, and image sequences to obtain the execution result of the localized image acquisition and processing flow.

[0026] Preferably, the step of performing local AI algorithm analysis and lesion identification processing on ultrasound image data acquired in offline mode, and generating an offline diagnostic report by combining the report template data in the local cache database, thus completing the offline business loop, includes:

[0027] Image feature vectors are extracted from ultrasound image data acquired offline, including grayscale histogram features, texture features, and shape features, to construct a multi-dimensional feature space.

[0028] A pre-trained deep learning model is applied to the multidimensional feature space to detect lesion regions. Abnormal tissue regions are identified through a convolutional neural network, and confidence scores are calculated to obtain the lesion region detection results.

[0029] Based on the detection results of the lesion area, quantitative measurement and analysis are performed to calculate geometric parameters such as lesion area, perimeter, major axis and minor axis, and generate structured measurement data.

[0030] The structured measurement data is matched and fused with the report template data in the local cache database to automatically fill in diagnostic parameters and suggested descriptions, generating an offline diagnostic report containing text, image annotations and AI analysis parameters;

[0031] The offline diagnostic report is subjected to integrity verification and format standardization processing to ensure compliance with medical report specifications, thus obtaining the completion status of the offline business loop.

[0032] Preferably, the continuous monitoring of the network connection status data, detection of network recovery events and triggering of the data synchronization process, identification of data to be synchronized during offline periods using a difference comparison algorithm, and execution of incremental data upload to the server include:

[0033] The data to be synchronized is constructed into a sparse graph structure, where each data unit is a node, and the dependencies and conflicts between data are connected by edges. The node weight value is calculated.

[0034] Based on the sparse graph structure, the maximum weighted independent set algorithm is applied to optimize the synchronization strategy. The distributed approximation algorithm is used to solve for the data subset that can be synchronized in parallel and has the largest weight.

[0035] The data subset is processed in batches according to the dependency relationship, and emergency data and basic dependency data with high weight values ​​are synchronized first to avoid data conflicts and transmission failures.

[0036] In a multi-client environment, a lightweight coordination protocol is used to coordinate synchronization plans, preventing server load surges and optimizing network resource utilization efficiency.

[0037] Perform status write-back and local cleanup operations on the synchronized data, mark the synchronized data and clear the temporary storage, and obtain the execution result of uploading the incremental data to the server.

[0038] Preferably, the feature dimensionality reduction processing based on the sparse tensor representation, using a local search principal component analysis algorithm, and finding the optimal principal component direction through gradient projection and soft thresholding operations, generates compressed feature data, including:

[0039] The sparse tensor representation is preprocessed to extract the positions and values ​​of non-zero elements in the medical image and construct a sparse matrix representation.

[0040] Based on the sparse matrix representation, an iterative optimization algorithm is applied to calculate the eigenvalues ​​and eigenvectors of the covariance matrix, and the principal component directions are found under the constraint of sparsity.

[0041] The principal component directions are subjected to soft thresholding. A threshold parameter is set, and coefficients smaller than the threshold parameter are set to zero to maintain sparsity, thus obtaining the principal component coefficients.

[0042] The principal component coefficients are updated by gradient projection method to ensure that the projection results satisfy the sparse constraint conditions. The optimization is iterated until convergence to obtain the optimized principal component coefficients.

[0043] Based on the optimized principal component coefficients, the compressed feature data is reconstructed, and the compression ratio and feature retention evaluation index are calculated to obtain the compressed feature data.

[0044] Preferably, the step of performing connected component analysis on the temporal correlation graph to identify missing frames, corrupted frames, and temporally anomalous regions in the image sequence, and to count the number and topological characteristics of connected components to obtain the connected component analysis results includes:

[0045] Based on the depth-first search algorithm, all nodes of the time series correlation graph are traversed to identify the set of interconnected nodes and obtain the connected components and the number of connected components.

[0046] For each connected component, topological characteristic parameters, including connectivity, diameter, clustering coefficient, and centrality index, are calculated to evaluate the integrity of the image sequence and obtain integrity results.

[0047] By analyzing the temporal span and spatial distribution characteristics of the connected components, the location of sequence breakage caused by equipment failure, transmission interruption, or buffer overflow can be identified.

[0048] The rationality of the connected components is verified based on the physiological periodicity pattern. For periodic images, the connected patterns corresponding to the physiological cycle are detected to obtain physiological verification results.

[0049] Based on the number of connected components, the integrity result, the sequence break location, and the physiological verification result, a connected component analysis result is constructed.

[0050] The connected component analysis results are compared with a preset integrity threshold. When the connectivity of any connected component is lower than the preset threshold, it is marked as an abnormal sequence that needs to be reconstructed, and a sequence quality assessment report is generated.

[0051] Preferably, the quantitative measurement and analysis based on the lesion area detection results, calculating geometric parameters such as lesion area, perimeter, major axis, and minor axis, and generating structured measurement data includes:

[0052] The detection results of the lesion area are processed by boundary extraction, and the lesion contour is identified by edge detection algorithm and morphological filtering is performed to remove noise.

[0053] Based on the lesion contour, the area parameters are calculated, the number of pixels in the lesion area is obtained by pixel statistics, and the actual area value is converted by combining the pixel spacing.

[0054] The perimeter of the lesion contour is measured, the perimeter of the lesion contour is calculated using a contour tracking algorithm, and a curve smoothing algorithm is applied to eliminate the jagged effect.

[0055] The major and minor axes of the lesion contour are calculated using the minimum bounding rectangle method to determine the main orientation and geometric parameters of the lesion contour.

[0056] Based on the geometric parameters, a shape descriptor is calculated, including morphological indices such as roundness, slenderness, and irregularity, to construct the structured measurement data containing location, size, shape, and texture features.

[0057] Preferably, the step of applying the maximum weighted independent set algorithm to optimize the synchronization strategy based on the sparse graph structure, and solving for the data subset with the largest weight that can be synchronized in parallel using a distributed approximation algorithm, includes:

[0058] The priority weights of the data to be synchronized are calculated using a multi-factor weighted calculation method, which includes data priority factor, time urgency factor and data size influence factor.

[0059] The graph is constructed based on data dependencies, including the association dependency between reports and images, the mapping dependency between patient IDs and examination data, and the conflict constraints of equipment resources.

[0060] A distributed algorithm is applied for multi-round iterative computation. In each round, each node updates its own state based on local information and coordinates global optimization through a message passing mechanism.

[0061] In each iteration, the candidate set of nodes is calculated, and the data item with the largest weight that does not conflict with the selected node is added to the independent set, until it can no longer be expanded;

[0062] The feasibility of the calculated maximum weighted independent set is verified to ensure that the selected data items meet the parallel synchronization conditions and do not violate the dependency constraints, and the optimal synchronization strategy scheme is generated.

[0063] Preferably, the compressed feature data is reconstructed based on the optimized principal component coefficients, and the compression ratio and feature retention evaluation index are calculated to obtain the compressed feature data, including:

[0064] Image data is reconstructed based on the principal component coefficients, and compressed features are mapped back to the original image space through matrix multiplication.

[0065] Calculate the compression ratio index and evaluate the degree of storage space saving and transmission efficiency improvement by comparing the data storage size before and after compression.

[0066] Assess the retention of key diagnostic information by comparing the contrast of lesion areas, edge sharpness, and the degree of preservation of texture details in images before and after compression.

[0067] Establish an image quality evaluation system, including objective evaluation indicators such as peak signal-to-noise ratio, structural similarity index, and visual information fidelity;

[0068] Based on the compression ratio index, the retention rate of key diagnostic information, and the objective evaluation index, an image quality index is constructed.

[0069] A quality threshold is set according to clinical diagnostic needs. When the image quality index is equal to or greater than the set quality threshold, the compression quality is confirmed to meet the diagnostic requirements, and the compressed feature data is generated.

[0070] Preferably, the step of comparing the connectivity analysis results with a preset integrity threshold, and marking any connected component with a connectivity score lower than the preset threshold as an abnormal sequence requiring reconstruction, and generating a sequence quality assessment report, includes:

[0071] Establish multi-level quality assessment standards, including pixel-level grayscale continuity assessment, feature-level anatomical structure consistency assessment, and semantic-level physiological parameter rationality assessment;

[0072] Pixel-level quality is quantitatively analyzed, and grayscale difference statistics, texture correlation coefficient and motion vector consistency index between adjacent frames are calculated.

[0073] Verify the spatial continuity of anatomical landmarks at the feature level and detect the geometric consistency of organ contours, blood vessel orientation, and tissue interfaces;

[0074] The clinical rationality of the sequence was verified by analyzing the trend of physiological parameter changes, including the examination of the reasonable range of physiological index changes and tissue movement amplitude.

[0075] Generate a sequence quality assessment report that includes quality score, abnormal location markers, and repair suggestions, providing clinicians with a basis for evaluating the reliability of imaging sequences and for diagnostic reference.

[0076] This invention also provides an offline ultrasound image acquisition and dynamic synchronous processing system, comprising:

[0077] The network status monitoring module is used to acquire network connection status data between the client and the server, continuously monitor the network connection status data and perform connectivity detection, determine the current network working mode, and obtain the network working mode result.

[0078] The online data synchronization module is used to retrieve pending inspection task data and report template data from the server when the network working mode result is online, store the pending inspection task data and the report template data in a local cache database, and establish an online data synchronization mechanism.

[0079] The image acquisition and processing module is used to acquire multimodal image data acquired by ultrasound equipment based on the online data synchronization mechanism, perform DICOM protocol parsing and format conversion processing on the multimodal image data to obtain converted image data, and write the converted image data to both a local temporary library and the server PACS system to generate local image backup data.

[0080] The offline mode switching module is used to detect network interruption events and trigger offline mode switching based on changes in the network connection status data, call the local SQLite database to take over the data storage service, activate the local image backup data, and start the local image acquisition and processing process.

[0081] The offline diagnostic processing module is used to perform local AI algorithm analysis and lesion identification processing on ultrasound image data acquired in offline state, and generate offline diagnostic reports by combining the report template data in the local cache database, thus completing the offline business closed loop.

[0082] The incremental synchronization module is used to continuously monitor the network connection status data, detect network recovery events and trigger the data synchronization process, identify the data to be synchronized generated during offline periods through a difference comparison algorithm, and perform incremental data upload to the server to complete the offline acquisition and dynamic synchronization processing of multimodal fusion ultrasound images.

[0083] The beneficial effects of this invention are:

[0084] 1. Improved the system's network adaptability and robustness, supporting inspection operations in both online and offline states;

[0085] 2. A network monitoring mechanism enables timely status awareness and ensures system availability through automatic reconnection and offline switching;

[0086] 3. Incremental synchronization + differential reconciliation mechanism to ensure data integrity;

[0087] 4. Significantly reduced operation and maintenance costs, with real-time performance monitoring and anomaly alerts;

[0088] 5. Automatic status recovery reduces fault recovery time by 80%, building a highly available, low-latency, and highly consistent data synchronization system. Attached Figure Description

[0089] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0090] Figure 1 A flowchart of the method for offline acquisition and dynamic synchronous processing of ultrasound images provided in an embodiment of the present invention;

[0091] Figure 2 This is a schematic diagram of the structure of the ultrasound image offline acquisition and dynamic synchronous processing system provided in an embodiment of the present invention. Detailed Implementation

[0092] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0093] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S1, S2, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0094] It will be understood by those skilled in the art that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application’s specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0095] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by those skilled in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0096] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0097] Example 1

[0098] like Figure 1 As shown, this embodiment provides a method for offline acquisition and dynamic synchronous processing of ultrasound images, including the following steps:

[0099] S11. Obtain network connection status data between the client and the server, continuously monitor and perform connectivity detection on the network connection status data, determine the current network working mode, and obtain the network working mode result.

[0100] Specifically, a network status monitoring module enables real-time monitoring of network connectivity. This module employs a combination of network detection mechanisms, including TCP connection testing, HTTP request response time analysis, and a heartbeat mechanism. The client sends a heartbeat to the server every preset time interval (e.g., 5 seconds) and records the response time. Simultaneously, the system periodically performs a deep connection test (e.g., every 30 seconds), including DNS resolution verification, route reachability testing, and service availability checks. Based on these detection results, the current network status is comprehensively assessed and categorized into three operating modes: "online," "weak network," or "offline." When three consecutive heartbeats fail to respond or an HTTP request times out, an additional verification mechanism is immediately triggered to confirm the authenticity of the network status change, avoiding misjudgments caused by network fluctuations and thus obtaining an accurate result for the network operating mode.

[0101] S12. When the network working mode result is online, retrieve the pending inspection task data and report template data from the server, store the pending inspection task data and the report template data in the local cache database, and establish an online data synchronization mechanism.

[0102] Specifically, a secure connection channel with the server is first established, employing OAuth 2.0 authentication and TLS 1.3 encryption to ensure secure data transmission. Based on the currently logged-in user's permissions and departmental assignments, the client retrieves a list of pending examination tasks from the server, including key data such as patient basic information (e.g., name, ID, age, gender), examination type (e.g., echocardiography, abdominal ultrasound), examination priority (e.g., routine, emergency, expedited), and appointment time. Simultaneously, the latest version of report template data is retrieved, including structured report templates, diagnostic terminology databases, measurement reference values, and AI-assisted diagnostic knowledge bases. This data uses an incremental synchronization strategy, transmitting only the parts that have changed since the last synchronization, significantly reducing data transmission volume. All retrieved data is stored in a local SQLite cache database, with a version control mechanism recording the last update time and synchronization status of each piece of data, laying the foundation for subsequent offline operations and data consistency management.

[0103] S13. Based on the online data synchronization mechanism, acquire multimodal image data collected by ultrasound equipment, perform DICOM protocol parsing and format conversion processing on the multimodal image data to obtain converted image data, and write the converted image data to both the local temporary library and the server PACS system to generate local image backup data.

[0104] Specifically, during the image acquisition phase, a connection is established with the ultrasound equipment via the DICOM protocol, supporting data acquisition for various ultrasound modes, including B-mode ultrasound (grayscale image), color Doppler, spectral Doppler, and elastography. The acquired raw DICOM data is first parsed to extract metadata such as patient information, equipment parameters, scan settings, and image attributes. Then, format standardization processing is performed, including pixel depth unification (converting to 16-bit grayscale), resolution adjustment, and compression ratio optimization. The processed image data is stored using a dual-write strategy: on one hand, the complete data is transferred to the hospital's PACS server via the DICOM Storage service, ensuring the data is integrated into the hospital's unified management system; on the other hand, a copy is stored in a local temporary repository, establishing a multi-level directory structure based on patient ID, examination type, and acquisition time, and generating a local index. The locally stored image data also undergoes preprocessing, including region of interest extraction, image enhancement, and keyframe marking, preparing for possible offline analysis, thereby generating fully functional local image backup data.

[0105] S14. Based on the changes in the network connection status data, detect network interruption events and trigger offline mode switching, call the local SQLite database to take over the data storage service, activate the local image backup data, and start the localized image acquisition and processing process.

[0106] Specifically, upon detecting a network interruption, the offline mode switching process is immediately initiated. First, a network interruption notification is sent to the user, but this does not affect the continuation of the current examination operation. Then, the data source is automatically switched, redirecting all data read and write operations from the remote server to the local SQLite database. At the database level, a transaction log recording mechanism is activated to record all data changes during the offline period in detail, providing a basis for subsequent data synchronization. For new image acquisition needs, a localized DICOM service is activated to establish a direct point-to-point connection with the ultrasound equipment, bypassing network dependencies. Simultaneously, previously cached local image backup data is activated, allowing medical staff to view historical examination results for reference. At the user interface level, the system displays a clear "Offline Mode" indicator and adjusts the function menu, retaining all core functions that can be performed offline, such as image acquisition, measurement analysis, and report editing, while disabling non-critical functions requiring a network connection, such as remote consultations, to ensure uninterrupted medical workflows.

[0107] S15. Perform local AI algorithm analysis and lesion identification processing on the ultrasound image data acquired in offline state, and generate an offline diagnostic report by combining the report template data in the local cache database, thus completing the offline business closed loop.

[0108] Specifically, in offline mode, pre-downloaded lightweight AI models are fully utilized to intelligently analyze newly acquired ultrasound images. These AI models are optimized to run efficiently with limited client computing resources, supporting the automatic identification and measurement of common lesions. Specifically, the images are first preprocessed, including noise suppression, contrast enhancement, and standardization; then, a multi-level convolutional neural network is applied for organ segmentation and lesion detection, identifying suspicious areas; next, fine classification and feature extraction are performed, calculating key parameters such as lesion size, shape, and echo characteristics; finally, structured analysis results are generated. Medical staff can review and modify this part of the AI-assisted analysis results and combine it with clinical experience for diagnosis. Locally cached report templates are invoked, automatically selecting the appropriate template based on the examination type, and automatically filling the template with patient information, examination parameters, measurement data, and AI analysis results. Doctors can edit diagnostic opinions and suggestions, add key image markers, and ultimately generate a standardized offline diagnostic report. After the report is completed, a local printing function is provided to ensure that patients can obtain their examination results in a timely manner, thus completing a full diagnostic loop from image acquisition to report generation.

[0109] S16. Continuously monitor the network connection status data, detect network recovery events and trigger the data synchronization process, identify the data to be synchronized generated during offline period through the difference comparison algorithm, perform incremental data upload to the server, and complete the offline acquisition and dynamic synchronization processing of multimodal fusion ultrasound images.

[0110] Specifically, in offline mode, the system continuously monitors network status in the background. Once network connectivity is restored, the data synchronization process is immediately initiated. First, a network stability assessment is performed to ensure a stable and reliable network connection before data synchronization begins, preventing interruptions due to network fluctuations. Then, the system reads the transaction logs from the offline period to identify all data items requiring synchronization, including newly added patient information, examination records, image data, and diagnostic reports. For each data type, the system employs different difference comparison algorithms: for structured data (such as patient information and examination records), a fast comparison based on timestamps and hash values ​​is used; for large binary data (such as image files), block hash comparison is used, transmitting only changed data blocks. An optimal synchronization strategy is formulated based on data priority and dependencies, typically executing synchronization in the order of "patient information → examination records → image data → diagnostic reports" to ensure data integrity and consistency. During synchronization, an intuitive progress display is provided, allowing users to continue working normally without affecting synchronization. For successfully synchronized data, the synchronization status flag in the local database is updated; for data that fails to synchronize due to conflicts or other reasons, detailed error reports and manual solutions are provided. This incremental and reliable data synchronization mechanism ensures seamless transfer of multimodal ultrasound imaging data between offline and online environments, minimizing the impact of network outages on medical work.

[0111] Example 2

[0112] Based on Example 1, the process involves acquiring multimodal image data from ultrasound equipment using the online data synchronization mechanism, performing DICOM protocol parsing and format conversion on the multimodal image data to obtain converted image data, and then dual-writing and storing the converted image data to a local temporary library and a server PACS system to generate local image backup data, including:

[0113] A three-dimensional sparse tensor structure is constructed for the multimodal image data, where the first dimension represents the spatial location coordinates, the second dimension represents the time series, and the third dimension represents the modal features. Regions of interest are extracted and sparse tensor representations are generated.

[0114] When processing multimodal ultrasound image data, the image data from different modalities (such as B-mode grayscale images, color Doppler, spectral Doppler, and elastography) are first organized into a unified three-dimensional sparse tensor structure. Specifically, the original DICOM data is first spatially registered to ensure the spatial correspondence of the same anatomical structures across different modalities; then, dynamic frames are arranged according to the acquisition time sequence to establish the temporal dimension; finally, the features of each modality are integrated as the third dimension. Since most areas in medical ultrasound images are background or low-information regions, an adaptive threshold segmentation algorithm is used to identify regions of interest (ROIs), retaining only diagnostically valuable tissue structure information, such as the ventricular wall and valve regions in cardiac ultrasound, and organ contours and lesion areas in abdominal ultrasound. For the identified ROIs, the system retains complete multimodal features; while background regions are highly compressed or directly zeroed out, thus forming a sparse representation. This sparse tensor structure not only significantly reduces data storage space requirements but also retains crucial diagnostic information, laying the foundation for subsequent feature extraction and analysis. For example, a typical echocardiogram might contain 60 frames of 512×512 pixel multimodal images, totaling about 150MB. However, by using sparse tensor representation, the amount of data can be reduced to 30-40% of the original, while retaining more than 95% of the diagnostic information.

[0115] Based on the sparse tensor representation, a local search principal component analysis algorithm is applied to perform feature dimensionality reduction. The optimal principal component direction is found through gradient projection and soft thresholding to generate compressed feature data.

[0116] In this step, an improved Local Search Principal Component Analysis (LSPCA) algorithm is used for feature dimensionality reduction, taking into account the characteristics of sparse tensors. Traditional PCA algorithms suffer from high computational complexity and are prone to losing local features when processing high-dimensional sparse data, while LSPCA overcomes these shortcomings through block processing and local feature preservation strategies. Specifically, the sparse tensor is first divided into multiple sub-regions according to anatomical structural features, such as the left ventricle, right ventricle, atrium, and valves in echocardiography. Then, principal component analysis is applied independently to each sub-region to extract local principal components. During principal component extraction, an iterative gradient projection method is used, employing alternating least squares and regularization constraints to solve for the optimal principal component direction. To maintain sparsity, a soft thresholding operation is applied to the calculated principal component coefficients, setting coefficients with absolute values ​​less than a preset threshold to zero. This ensures compression while avoiding the loss of detail due to excessive smoothing. Through this localized feature dimensionality reduction, the system can capture key diagnostic features in different regions, such as myocardial motion patterns, hemodynamic characteristics, and tissue elasticity changes. Finally, the principal components and their coefficients of each sub-region are recombined to form complete compressed feature data, further reducing the data volume to 15-25% of the original data, creating conditions for offline storage and processing.

[0117] The compressed feature data is quality assessed, and the variance explanation rate and the retention rate of key diagnostic information are calculated. When the variance explanation rate reaches a preset validity threshold, the compression validity is confirmed, and the compressed image data is stored in the local temporary library.

[0118] To ensure that the compressed image data still retains sufficient diagnostic value, a rigorous quality assessment mechanism was established. First, the variance explained rate of the compressed feature data was calculated, i.e., the proportion of the original data's total variance that the retained principal components could explain. For different types of ultrasound examinations, the system set differentiated variance explained rate thresholds: for examinations requiring detailed structural analysis (such as heart valve assessment), a variance explained rate of no less than 92% was required; for examinations primarily focused on gross morphology (such as routine abdominal screening), the threshold could be appropriately lowered to 85%. In addition to statistical indicators, the system also assessed the retention of key diagnostic information, including three aspects: structural integrity score, comparing the differences in anatomical boundaries before and after compression using edge detection and contour matching algorithms; texture fidelity score, comparing the retention of tissue texture features using Local Binary Pattern (LBP) and Gray-Level Co-occurrence Matrix (GLCM) features; and dynamic characteristics score, assessing the fidelity of motion trajectories and periodic changes for dynamic organs such as the heart. These scores were then considered in conjunction with the variance explained rate. When the overall score exceeded a preset threshold, the compression was confirmed as effective, and the compressed image data was organized and stored in a local temporary library according to patient ID, examination type, and timestamp. For data that fails the quality assessment, the compression parameters will be automatically adjusted, such as increasing the number of principal components to be retained or lowering the soft threshold parameter. Then the compression and assessment process will be re-executed until the quality requirements are met or the maximum number of attempts is reached.

[0119] Based on the compressed feature data, image statistical features, texture descriptors, and morphological parameters are calculated to generate a preprocessed feature library required for offline AI analysis, thus obtaining the local image backup data.

[0120] After image compression and storage, rich image features are further extracted to construct a comprehensive preprocessed feature library, supporting AI-assisted diagnosis in offline mode. First, the system calculates basic statistical features, including first- and second-order statistics such as the gray-level distribution histogram, mean, variance, skewness, and kurtosis of each region of interest. These features reflect tissue echo characteristics and homogeneity. Second, multi-scale texture descriptors are extracted, including Gabor filter features, wavelet transform coefficients, local binary patterns, and gray-level co-occurrence matrix derived features (such as contrast, correlation, energy, and homogeneity). These texture features are valuable for identifying different types of tissues and lesions. Third, morphological parameters are calculated, including region area, perimeter, roundness, eccentricity, and orientation. These parameters are crucial for describing organ morphology and lesion characteristics. For dynamic sequences, temporal features are also extracted, such as motion trajectories, velocity curves, and periodic change patterns. These features are particularly important for cardiac function assessment and hemodynamic analysis. All extracted features are standardized and organized according to feature type and anatomical region to form a structured feature vector set. These preprocessed features, along with the compressed image data, constitute a complete local image backup, stored in a local database. When entering offline mode, the AI ​​diagnostic module can directly access these preprocessed features without re-extracting them from the original images, greatly improving analysis efficiency. Even on client devices with limited computing resources, complex lesion identification and measurement functions can be achieved.

[0121] Example 3

[0122] Based on Example 1, the step of detecting network interruption events and triggering offline mode switching based on changes in network connection status data, calling the local SQLite database to take over data storage services, activating the local image backup data, and starting the localized image acquisition and processing flow includes:

[0123] Each frame in the dynamic ultrasound image sequence acquired offline is constructed as a node in a graph structure, and the spatiotemporal correlation between frames is calculated as the edge weight to establish a temporal correlation graph.

[0124] In offline mode, it is crucial to ensure the integrity and continuity of dynamic ultrasound image sequences, especially when network or equipment failures may lead to frame loss or corruption. To address this, a graph theory modeling approach is employed, constructing each ultrasound image frame as a node in a graph structure and establishing weighted connections by calculating inter-frame correlations. Specifically, key feature points are first extracted from each frame, such as endocardial boundary points, valve locations, and vascular wall landmarks in cardiac ultrasound. Algorithms like SIFT or ORB are used to ensure these feature points are rotation- and scaling-invariant. Then, the matching degree and displacement vectors of feature points between adjacent frames are calculated to generate a spatiotemporal correlation index. This index comprehensively considers feature point matching rate, motion consistency, and image structural similarity, forming a normalized edge weight value between 0 and 1. For organs with periodic motion (such as the heart), the system also establishes cross-cycle connections, creating weak connections (typically 30-50% of the weight of adjacent frame connections) between frames of the same phase in different cardiac cycles, forming a temporal correlation graph with a periodic structure. This graph structure not only reflects the temporal continuity of the image sequence but also contains physiological cycle information, providing a theoretical basis for subsequent sequence integrity analysis and repair. For example, in a typical echocardiogram, a sequence containing 60 frames will form a weighted undirected graph with approximately 120 edges (including adjacent frame connections and periodic connections), whose topological structure intuitively reflects the integrity of the cardiac cycle.

[0125] Connectivity component analysis is performed on the temporal correlation graph to identify missing frames, damaged frames, and temporal anomaly regions in the image sequence. The number of connected components and topological characteristics are counted to obtain the connectivity component analysis results.

[0126] Based on the constructed temporal correlation graph, graph theory algorithms are applied for in-depth connectivity analysis to assess the integrity of the image sequence and identify potential problems. First, an edge weight threshold (typically 0.4-0.6) is set, and weak connections below the threshold are temporarily removed. Then, depth-first search or breadth-first search algorithms are applied to identify all connected components in the graph. Ideally, a complete image sequence should form a single connected component; the presence of multiple connected components indicates a break in the sequence. The system further calculates the topological properties of each connected component, including the number of nodes (reflecting subsequence length), average degree (reflecting the tightness of inter-frame connections), clustering coefficient (reflecting local connectivity patterns), and diameter (reflecting the distance between the farthest nodes). By analyzing these topological parameters, problem regions in the sequence can be precisely located: isolated single-frame nodes usually indicate damaged frames; "missing bridge points" between connected components indicate missing frames; and low-weight regions within connected components may be temporally abnormal regions. For periodic sequences (such as echocardiography), the periodic characteristics of connected components are also analyzed to detect the existence of a complete periodic structure and the consistency between periods. All these analysis results are integrated into a structured connected component analysis report, which includes a sequence integrity score (0-100 points), problem region location (frame index range), and suggested repair strategies. For example, it may identify specific problems such as "frames 15-18 are missing, resulting in an incomplete first cardiac cycle" or "frame 33 is corrupted and needs to be replaced," providing precise guidance for subsequent sequence repair.

[0127] Based on the connected component analysis results and combined with the preprocessed feature library in the local image backup data, the optimal graph reconstruction algorithm is applied to repair the integrity of the image sequence. Missing data fragments are filled in by feature interpolation and periodic reconstruction methods to obtain the reconstructed image sequence.

[0128] After identifying the problem areas in the image sequence, an intelligent restoration process is initiated, employing differentiated reconstruction strategies for different types of problems. For missing frames, the system first attempts to find possible redundant backups in local image backup data; if no backup is available, optimal interpolation algorithms are applied based on graph reconstruction theory. Specifically, feature vectors of the frames before and after the missing frame are extracted from the preprocessed feature library to construct a feature space trajectory, and then feature vectors at the missing location are generated through spline interpolation or Bézier curve fitting. For periodic sequences (such as echocardiography), periodic reconstruction methods are prioritized, i.e., frames of the corresponding phases from other complete cycles are found as references, and adaptive adjustments are made in conjunction with local deformation models. For example, if frames 15-18 of the first cardiac cycle are missing, the corresponding frames from the second cycle are referenced, and corrections are made based on the overall difference between the two cycles. For damaged frames, an image-based restoration method is used, preserving the effective regions in the damaged frames and reconstructing only the damaged areas. During the restoration process, statistical features and texture descriptors in the preprocessed feature library are fully utilized to ensure that the reconstructed content is consistent with the original tissue characteristics. For all reconstructed frames, a confidence score is calculated, and their reconstruction status is marked in the metadata so that doctors can understand the data source during diagnosis. This intelligent sequence restoration process maximizes the restoration of the integrity and continuity of image sequences while ensuring clinical effectiveness, thus improving the reliability of offline diagnosis.

[0129] The reconstructed image sequence is subjected to multi-level consistency verification, including pixel-level grayscale continuity verification, feature-level anatomical structure continuity verification, and semantic-level physiological parameter rationality verification, to generate complete localized image data.

[0130] To ensure that the reconstructed image sequences meet clinical diagnostic requirements, a rigorous multi-level consistency verification mechanism was established. In pixel-level verification, statistical measures of gray-level differences between adjacent frames are calculated, including objective indicators such as mean absolute difference (MAD), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR), ensuring smooth and natural transitions between frames without significant brightness jumps or noise abrupt changes. The system also analyzes the changing trends of the gray-level co-occurrence matrix to verify the continuity of texture features. In feature-level verification, the motion trajectories of key anatomical structures are tracked, such as the endocardial boundary, valve opening and closing positions, and vascular wall motion in echocardiography. By calculating the displacement vectors and deformation parameters of these structures in consecutive frames, abnormal motion patterns that do not conform to physiological laws can be detected. In particular, the motion patterns of reconstructed frames are compared with those of the original frames to ensure that the reconstructed content does not introduce unnatural motion artifacts. In semantic-level verification, the time series of key physiological parameters are calculated, such as changes in ventricular wall thickness, chamber area, and blood flow velocity curves in echocardiography, and compared with standard physiological models to ensure that the changing trends of these parameters conform to medical principles. For example, the ventricular wall should thicken rather than thin during cardiac systole, and the blood flow velocity profile should exhibit specific waveform characteristics. Regions that fail validation are marked as "low confidence areas" and provided with detailed problem descriptions, such as "discontinuous ventricular motion trajectory in frames 20-22" or "abnormally high blood flow velocity in frame 45." This comprehensive consistency validation generates complete and reliable localized imaging data, providing a solid foundation for offline diagnosis.

[0131] The complete localized image data is stored hierarchically according to patient ID, examination items, and image sequences to obtain the execution result of the localized image acquisition and processing flow.

[0132] After image sequence reconstruction and validation, a structured hierarchical storage strategy is adopted to ensure orderly data organization and efficient access. At the top level, patient ID serves as the primary index, with each patient possessing a unique identifier in the local database. For newly added patients offline, temporary IDs are generated and key demographic information is recorded to prepare for subsequent synchronization with the hospital's master index (EMPI). At the second level, data is categorized by examination item, such as cardiac ultrasound, abdominal ultrasound, and vascular ultrasound. Each examination item includes metadata such as examination time, administering physician, and examination purpose. At the third level, specific image sequences are stored, such as four-chamber views, long-axis views, and short-axis views in cardiac ultrasound. Each sequence contains a complete set of frames, sequence description, and technical parameters. For reconstructed or repaired sequences, special markers are added to the metadata, detailing the repaired area, the method used, and the confidence score to ensure data traceability. In addition to the original DICOM data, derived data, such as preprocessing features, measurement results, and AI analysis outputs, are also stored, maintaining a strict correlation with the original images. To optimize storage efficiency, a tiered storage strategy is adopted: frequently accessed data (such as data from the current day's examinations) is stored in a cache; historical data is compressed or archived based on access frequency and clinical importance. All storage operations are recorded in a transaction log, including timestamps, operation types, and data identifiers, providing a basis for subsequent network recovery and synchronization. This structured, tiered storage mechanism not only ensures the integrity and accessibility of offline data but also lays the foundation for subsequent online synchronization, ultimately forming a complete localized image acquisition and processing workflow.

[0133] Example 4

[0134] Based on Example 1, the process of performing local AI algorithm analysis and lesion identification on ultrasound image data acquired in offline mode, and generating an offline diagnostic report by combining the report template data in the local cache database, thus completing the offline business loop, includes:

[0135] Image feature vectors are extracted from ultrasound image data acquired offline, including grayscale histogram features, texture features, and shape features, to construct a multidimensional feature space.

[0136] In an offline environment with a disconnected network, comprehensive feature extraction of the acquired ultrasound images is necessary to lay the foundation for subsequent lesion detection and analysis. A multi-level feature extraction strategy is employed, systematically extracting features from low-level statistical features to high-level semantic features. At the gray-level histogram feature level, first-order statistics are calculated, including mean gray value, standard deviation, skewness, kurtosis, energy, and entropy. These features reflect the overall distribution characteristics of tissue echoes. Gray-level quantile features (such as the 10th, 25th, 50th, 75th, and 90th percentile values) are also calculated to more comprehensively describe the gray-level distribution. For the unique echo characteristics of ultrasound, the proportion of low-echo areas, the distribution of strong echo points, and echo uniformity indices are specifically extracted. At the texture feature level, several complementary texture description algorithms were implemented: features based on the Gray-Level Co-occurrence Matrix (GLCM), including contrast, correlation, energy, homogeneity, and entropy, are calculated at different distances and directions to capture the directional texture features of tissue; features based on Local Binary Patterns (LBP), this rotation-invariant texture descriptor is particularly suitable for speckle pattern analysis in ultrasound images; features based on wavelet transform, capturing texture information of different frequencies through multi-scale decomposition; and features based on fractal dimension, quantifying the complexity and roughness of the tissue surface. At the shape feature level, the contour of the region of interest (ROI) is first extracted using adaptive thresholding or region growing algorithms, and then the shape descriptor is calculated: basic geometric features such as area, perimeter, roundness, rectangularity, aspect ratio, and compactness; moment invariants, providing rotation-, scaling-, and translation-invariant shape descriptions; Fourier descriptors, capturing detailed changes in the contour; and shape context features, describing the relative distribution of boundary points. Edge features are also extracted, including edge sharpness, edge gradient direction consistency, and edge integrity, which are particularly important for distinguishing between benign and malignant lesions. All extracted features are normalized to form high-dimensional feature vectors, creating a multi-dimensional feature space describing the content of the ultrasound image. Principal component analysis (PCA) or t-SNE is used to project the high-dimensional features into a low-dimensional space, facilitating visualization and analysis and reducing computational complexity. This comprehensive feature extraction provides a rich information foundation for subsequent lesion detection and classification, ensuring accuracy and stability even in offline environments.

[0137] A pre-trained deep learning model is applied to the multidimensional feature space to detect lesion regions. Abnormal tissue regions are identified through a convolutional neural network, and confidence scores are calculated to obtain the lesion region detection results.

[0138] After feature extraction, the system uses a locally cached pre-trained deep learning model to automatically detect and analyze lesion regions in ultrasound images. A multi-stage detection strategy is employed, combining traditional and deep learning features to ensure stability and accuracy in offline environments. First, the system uses a lightweight anomaly region candidate generation network, based on U-Net or Mask R-CNN architectures, to quickly filter regions that may contain lesions. These models are optimized with parameter counts controlled within the 5-10M range, ensuring efficient operation even on portable devices. Candidate region generation employs sliding window or Region Proposal Network (RPN) techniques, optimized for the characteristics of ultrasound images, such as special processing for acoustic attenuation and acoustic shadowing features. Second, the system applies a refined classification network to each candidate region. These networks are typically based on ResNet, DenseNet, or EfficientNet architectures, but compressed using knowledge distillation techniques to maintain high accuracy while reducing model size. The classification network outputs a lesion type prediction (e.g., cyst, solid mass, calcification) and a malignancy risk score (0-100 points). The system employs an ensemble learning strategy, fusing predictions from multiple pre-trained models to improve detection stability. For each detected lesion, a multidimensional confidence score is calculated: detection confidence, reflecting the probability that the region contains a lesion; classification confidence, reflecting the reliability of lesion type judgment; and malignancy confidence, reflecting the certainty of malignancy risk assessment. An attention mechanism is also implemented to generate a heatmap displaying the image region that contributes most to the classification decision, improving the interpretability of the results. For ultrasound-specific lesion types, dedicated detection modules are integrated: for breast ultrasound, BI-RADS features such as morphology, boundary, and posterior echo are specifically detected; for thyroid ultrasound, TI-RADS features such as echo, boundary, and calcification are specifically analyzed; and for liver ultrasound, LI-RADS features such as enhancement pattern and capsule are specifically evaluated. The system generates a structured detection result report, including lesion location (coordinates and anatomical localization), size, type, malignancy risk score, feature description, and confidence index. For detection results with a confidence score below a threshold, the system marks them as "recommended for manual review" to ensure clinical safety. This multi-layered lesion detection strategy can provide accurate and reliable auxiliary diagnostic information even in offline environments.

[0139] Based on the detection results of the lesion area, quantitative measurement and analysis are performed to calculate geometric parameters such as lesion area, perimeter, major axis and minor axis, and generate structured measurement data.

[0140] After successfully detecting the lesion area, the system enters the precise quantitative measurement stage, providing objective and repeatable numerical references for clinical diagnosis. First, the detected lesion area is finely segmented, and precise boundaries are extracted using deep learning segmentation models (such as DeepLab and SegNet) or traditional image processing methods (such as level set methods and region growing algorithms). To improve segmentation accuracy, the system implements an interactive correction mechanism, allowing operators to adjust the automatic segmentation results when necessary. Based on the segmentation results, comprehensive geometric parameters are calculated: basic dimensional parameters, including area (square millimeters), perimeter (millimeters), maximum diameter (millimeters), minor axis perpendicular to the maximum diameter (millimeters), and average diameter; shape parameters, including roundness, aspect ratio (major axis / minor axis), compactness (4π × area / perimeter²), and irregularity (the ratio of the actual perimeter to the circumference of a circle of equal area); and orientation parameters, including principal axis direction (angle relative to the horizontal axis) and orientation entropy (uniformity of boundary orientation distribution). For three-dimensional ultrasound data, the system also calculates volume (cubic millimeters), surface area (square millimeters), and sphericity (36π × volume² / surface area³). Special attention is paid to clinically significant characteristic measurements: for breast masses, the anteroposterior diameter ratio and skin distance are measured; for thyroid nodules, the anteroposterior diameter / transverse diameter ratio and capsule distance are measured; for liver lesions, the echogenicity ratio of the parenchyma to the surrounding tissue is measured. Multi-timepoint measurement comparison functionality is also implemented. When historical examination data is available in the local database, the lesion size change rate and volume doubling time are automatically calculated—parameters crucial for assessing lesion growth rate. All measurement results undergo rigorous quality control, including measurement uncertainty assessment and reliability labeling. Structured measurement data packages are generated, and all parameters are stored in standardized formats (such as DICOM SR or HL7 FHIR) to ensure data interoperability and traceability. Measurement data also includes metadata information, such as measurement time, algorithm version used, operator identification, and calibration information, improving data reliability. This comprehensive and accurate quantitative measurement provides an objective basis for clinical decision-making, especially valuable in cases requiring long-term follow-up comparisons.

[0141] The structured measurement data is matched and fused with the report template data in the local cache database to automatically fill in diagnostic parameters and suggested descriptions, generating an offline diagnostic report containing text, image annotations, and AI analysis parameters.

[0142] After quantitative measurement is completed, the report generation stage begins, transforming the analysis results into standardized and professional medical diagnostic reports. First, applicable report templates are retrieved from the local cache database. The most suitable template is selected based on the examination type (e.g., breast ultrasound, thyroid ultrasound, abdominal ultrasound), clinical purpose (e.g., initial diagnosis, follow-up, screening), and lesion characteristics (e.g., solid, cystic, mixed). The report templates employ a structured design, including fixed text paragraphs and variable parameter placeholders, supporting conditional logic and branching structures. Structured measurement data is intelligently integrated with the templates: numerical parameters are directly filled into corresponding placeholders, such as "maximum diameter of lesion {long diameter} mm"; classification results are converted into standard descriptive terms, such as converting echo pattern classification results into "hypoechoic," "isoechoic," or "hyperechoic"; and measurement parameter combinations generate composite descriptions, such as generating descriptions of "clear and regular boundaries" or "fuzzy and irregular boundaries" based on shape parameter combinations. A rule-based diagnostic suggestion generation mechanism is implemented, retrieving matching diagnostic and follow-up suggestions from the knowledge base based on lesion characteristics, measurement parameters, and risk scores. For example, for breast lesions, classifications and recommendations are automatically generated based on the BI-RADS scoring criteria; for thyroid nodules, treatment recommendations are provided based on the TI-RADS scoring criteria. Image annotation is also integrated, automatically marking lesion locations, outlines, and measurement lines on key images, generating annotated report images. Annotations use standardized visual coding: different colors are used for different types of lesions (e.g., red for high malignancy risk, yellow for intermediate risk, and green for low risk); measurement lines use standard symbols (e.g., solid lines for the long axis and dashed lines for the short axis); confidence levels are represented by line thickness or transparency. AI analysis parameters are integrated into the report, including detection confidence levels, risk scores, and key feature assessments, but these are explicitly labeled as "AI-assisted analysis results" to avoid confusion with physician diagnoses. The report also includes a methodological description section, briefly describing the AI ​​model used, its version number, and its scope of application, enhancing the report's scientific rigor and transparency. The final offline diagnostic report adopts a hierarchical structure: the summary section concisely lists key findings and recommendations; the detailed description section includes complete imaging features, measurement results, and analysis; the image section displays labeled key images; and the appendix contains complete measurement data and AI analysis parameters. This structured and intelligent report generation mechanism can provide professional and standardized diagnostic reports even in offline environments, improving diagnostic efficiency and quality.

[0143] The offline diagnostic report is subjected to integrity verification and format standardization processing to ensure compliance with medical report specifications, thus obtaining the completion status of the offline business loop.

[0144] After generating the initial report, the final quality control and standardization phase begins, ensuring that the offline-generated report conforms to medical guidelines and institutional standards. First, a comprehensive integrity check is performed to verify that all necessary components of the report are complete: patient basic information, including name, ID, age, gender, and examination date; examination information, including examination type, equipment used, and examining physician; description of imaging findings, including normal and abnormal findings; measurement data, including quantitative parameters for all lesions; diagnostic impressions, including summative assessment and classification; and recommendations, including follow-up or further examination suggestions. The system checks the completeness of each section, marking missing or incomplete fields, such as incomplete measurement data, insufficient descriptions, or missing recommendations. Second, a language standardization check is performed to ensure the report text conforms to medical expression standards: terminology consistency check ensures the use of standardized medical terminology, avoiding different expressions for the same concept; grammar and spelling check corrects possible grammatical and spelling errors; and expression integrity check ensures clear and logically coherent descriptions, avoiding vague or ambiguous expressions. The standardization of numerical representation is also checked, such as consistent units of measurement, appropriate decimal places, and reasonable numerical ranges. Third, the system performs format standardization to ensure report layout conforms to institutional standards: applying predefined format templates, unifying fonts, font sizes, line spacing, and margins; standardizing headers and footers, including institutional logos, page numbers, and report IDs; standardizing image layouts, ensuring appropriate image sizes and clear labeling; and generating PDF versions to ensure consistent display across different systems. It also performs medical compliance checks to ensure reports comply with relevant regulations and guidelines: privacy protection checks to ensure no unnecessary sensitive information is included; disclaimer checks to ensure AI-assisted content is appropriately labeled and explained; and compliance statement checks to ensure necessary legal and ethical statements are included. Finally, report metadata is generated, recording key information from the report generation process: report version number and generation time; AI model version and parameters used; operator identity and operation logs; integrity check results and correction records. The completed report is stored in a local database, marked as "offline generation - pending synchronization," awaiting network recovery before being uploaded to the central system. Simultaneously, the local business status is updated, marking the current check as "offline closed-loop completion," and recording the completion time and report ID. Through this series of rigorous quality control and standardization processes, we ensure that even diagnostic reports generated in an offline environment meet the same professional standards and specifications as those in an online environment, truly achieving a complete closed loop for offline business.

[0145] Example 5

[0146] Based on Example 1, the continuous monitoring of the network connection status data, detection of network recovery events and triggering of the data synchronization process, identification of data to be synchronized during offline periods using a difference comparison algorithm, and execution of incremental data upload to the server include:

[0147] The data to be synchronized is constructed into a sparse graph structure, where each data unit is a node, and the dependencies and conflicts between data are connected by edges. The node weight values ​​are calculated.

[0148] Once the network is restored, the primary challenge is efficiently synchronizing the large amount of heterogeneous data generated during the offline period. To address this, a graph theory modeling approach is employed, constructing the data to be synchronized as a sparse graph structure. In this structure, each data unit (such as patient information records, examination tasks, DICOM image files, diagnostic reports, etc.) is represented as a node in the graph. Two types of edges are established based on the logical relationships between the data: dependency edges and conflict edges. Dependency edges represent the order constraints between data, such as basic patient information must be synchronized before the patient's examination records, and image data must be synchronized before reports that reference these images. Conflict edges represent data pairs that cannot be processed in parallel, such as large DICOM files generated by the same ultrasound device not being suitable for parallel transmission (potentially leading to device I / O contention), and the risk of overwriting between multiple examination status updates for the same patient. For each node, a comprehensive weight value is calculated using a multi-factor weighted formula: W = α×P + β×T + γ×S, where P represents data priority (emergency = 10, routine = 5, physical examination = 1), T represents time urgency (using the reciprocal function of the current time minus the creation time), and S represents the data size impact factor (small files have slightly higher weight than large files for faster synchronization). α, β, and γ are adjustable weight coefficients, optimized according to the hospital's business characteristics. For example, in the emergency department, the α value can be increased to prioritize the synchronization of emergency data; in environments with limited network bandwidth, the γ value can be increased to prioritize the transmission of small files. This graph structure not only intuitively expresses the complex relationships between data but also provides a theoretical basis for subsequent synchronization strategy optimization.

[0149] Based on the sparse graph structure, the maximum weighted independent set algorithm is applied to optimize the synchronization strategy. A distributed approximation algorithm is used to solve for the data subset with the largest weight that can be synchronized in parallel.

[0150] After constructing the sparse graph structure, the core problem becomes: how to select the data set with the largest sum of weights for parallel synchronization while satisfying dependency constraints and avoiding conflicts. This is essentially the "maximum weighted independent set" problem in graph theory, which involves selecting a set of non-adjacent nodes in a graph such that the sum of their weights is maximized. Since this problem is NP-hard, the system employs an improved distributed approximation algorithm to solve it. This algorithm is based on the Congest model and is suitable for execution in distributed environments with limited communication bandwidth. The algorithm execution process consists of multiple iterations: first, each node broadcasts its own weight and neighbor node information to other nodes within a certain range; then, each node calculates its local decision value (the ratio of its weight to the number of neighbor nodes) based on the received information; next, the node with the highest decision value is added to the independent set, and it and all its neighbors are marked as "processed"; finally, the above steps are repeated for the remaining unprocessed nodes until all nodes are processed. To improve algorithm efficiency, preprocessing optimization is introduced: for explicit linear dependency chains (such as patient information → examination records → image data → report), the system pre-calculates the optimal transmission sequence to reduce algorithm runtime. The algorithm also employs an adaptive parameter adjustment mechanism, dynamically adjusting the iteration step size and termination conditions based on network conditions and server load. Practice shows that the algorithm can find a near-optimal solution (typically reaching 85-95% of the theoretical optimal solution) within a near-linear time complexity, meeting the real-time requirements of medical data synchronization. Through this intelligent optimization, the system can find the optimal parallel synchronization strategy in complex data-dependent networks, significantly improving data synchronization efficiency.

[0151] The data subsets are processed in batches according to their dependencies, with priority given to synchronizing emergency data and basic dependency data with high weight values ​​to avoid data conflicts and transmission failures.

[0152] After obtaining the maximum weighted independent set, the system further refines the scheduling and execution control of synchronization tasks. First, the system arranges the nodes in the independent set in descending order of their weights, forming an initial priority queue. Then, a topological sorting algorithm is applied to ensure that all dependency constraints are met, meaning that all dependencies of a data item must be synchronized before that data item. Specifically, "basic dependent data," i.e., key nodes depended on by multiple high-weight data (such as patient master index records), are identified and their priority is increased to ensure that these data are synchronized first, paving the way for subsequent high-weight data synchronization. For emergency data (such as images and reports from emergency ultrasound examinations), a dedicated fast channel is set up, allowing these data to start synchronization immediately after meeting the minimum dependency conditions, without waiting for all peer dependencies to complete. An intelligent batch partitioning mechanism is also implemented, dividing synchronization tasks into multiple batches. Tasks within each batch can be executed in parallel, while batches are executed strictly in sequence. Batch partitioning considers factors such as network bandwidth, server processing capacity, and data characteristics to avoid network congestion or server overload caused by excessive concurrency. For large data files (such as high-resolution ultrasound image sequences), a chunked transmission strategy is applied, splitting a single large file into multiple data blocks for parallel transmission. Simultaneously, verification and error recovery mechanisms are implemented to ensure data integrity. The system also establishes a real-time monitoring and adaptive adjustment mechanism to continuously monitor synchronization progress and network conditions. When transmission anomalies are detected (such as a sudden drop in speed or an increase in error rate), the system automatically adjusts the concurrency and retry strategy to maximize the successful completion of synchronization tasks.

[0153] In a multi-client environment, a lightweight coordination protocol is used to coordinate synchronization plans, preventing server load surges and optimizing network resource utilization efficiency.

[0154] In large medical facility environments, multiple ultrasound workstations typically operate simultaneously. When the network recovers, if all clients initiate data synchronization at the same time, it can lead to a surge in server load and network congestion. To address this issue, a distributed synchronization coordination mechanism based on a lightweight coordination protocol was implemented. This protocol employs a two-layer architecture: the server layer is responsible for global resource allocation and load balancing; the client layer is responsible for local synchronization strategy optimization and execution. Upon network recovery, the client first sends a synchronization request to the server, containing a summary of the data to be synchronized (data type, quantity, total size, and priority distribution). Based on the current system load and the requests from all clients, the server calculates the optimal synchronization time window allocation scheme, assigning each client a specific synchronization period and resource quotas (such as bandwidth limits and maximum concurrent connections). After receiving the allocation results, the client executes the synchronization task within the specified time window and reports the synchronization progress to the server in real time. The server continuously monitors the overall synchronization status, dynamically adjusting the resource allocation for other clients when a client completes synchronization or exits prematurely, ensuring efficient utilization of system resources. To address network fluctuations, the protocol implements a heartbeat mechanism and session recovery functionality. When reconnecting after a brief network interruption, synchronization can resume from the point of failure, avoiding duplicate transmissions. The protocol also supports priority preemption; when a high-priority client (such as an emergency room workstation) requests synchronization, the server can temporarily reduce the resource quota for lower-priority clients, ensuring that critical data is processed first. This coordination mechanism effectively avoids the "synchronization storm" problem, keeping server load within a reasonable range while ensuring fair and efficient allocation of network resources, significantly improving overall synchronization efficiency and system stability.

[0155] Perform status write-back and local cleanup operations on the synchronized data, mark the synchronized data and clear the temporary storage, and obtain the execution result of uploading the incremental data to the server.

[0156] After the data synchronization task is completed, the final state integration and cleanup phase begins to ensure data consistency and storage efficiency. First, for each successfully synchronized data item, the server returns confirmation information, including the server's unique identifier, storage location, and version information. The client updates the synchronization status flag in its local database based on this information, changing the data item's status from "pending synchronization" to "synchronized," and recording the synchronization completion time and the identifier assigned by the server. For temporary patient IDs created during offline periods, a mapping relationship is established with the hospital's master index (EMPI) to ensure that these data can be accessed subsequently via official IDs. For data items that fail to synchronize, the system records detailed error information and the reason for the failure, and formulates different handling strategies based on the error type: for failures caused by temporary network failures, automatic retry is arranged; for failures caused by data conflicts, a conflict report is generated and manual resolution options are provided; for failures caused by permission issues, a permission request process is triggered. After the status update is complete, storage space optimization is performed: temporary data that has been successfully synchronized and is no longer needed locally (such as intermediate processing results and temporary cache files) is safely deleted; for data that needs to be retained locally (such as recent test results), its access frequency and clinical importance are assessed to determine whether to compress storage or move it to a low-speed storage area. The system also generates a detailed synchronization report, including statistics on synchronized data (such as success rate, average speed, and total time), a summary of anomalies, and changes in storage space, providing administrators with a comprehensive monitoring view. Finally, the synchronization status is reset, restoring normal online working mode and starting a new round of real-time data synchronization. Through these comprehensive status write-back and cleanup operations, seamless integration of offline data with the online system is ensured, while local storage resources are optimized, providing a clear and consistent data environment for subsequent business operations.

[0157] Example 6

[0158] Based on Example 2, the feature dimensionality reduction process is performed using the local search principal component analysis algorithm based on the sparse tensor representation. The optimal principal component direction is found through gradient projection and soft thresholding to generate compressed feature data, including:

[0159] The sparse tensor representation is preprocessed to extract the positions and values ​​of non-zero elements in the medical image and construct a sparse matrix representation.

[0160] Before performing principal component analysis, the three-dimensional sparse tensor needs to be preprocessed and transformed to meet the requirements of subsequent algorithms. The preprocessing stage first uses a scanning algorithm to identify and extract non-zero elements from the tensor, recording the three-dimensional coordinates (i,j,k) and corresponding value v of each non-zero element, where i represents the spatial coordinates, j represents the time series index, and k represents the modal feature dimension. Due to the characteristics of medical ultrasound images, non-zero elements are usually concentrated in the region of interest (ROI), such as the heart chambers and myocardium in cardiac ultrasound, accounting for 30-40% of the overall data volume. Therefore, this representation significantly reduces storage and computational overhead. Next, the system uses tensor unfolding technology to convert the three-dimensional tensor into a two-dimensional sparse matrix representation. Specifically, the system selects the unfolding method most suitable for the current analysis task: for analyses focusing on spatial features, mode-1 unfolding is used, with the spatial dimension as rows and the combination of time and modal dimensions as columns; for analyses focusing on temporal features, mode-2 unfolding is used, with the time dimension as rows and the combination of spatial and modal dimensions as columns. The expanded sparse matrix is ​​efficiently stored using Compressed Row Storage (CSR) or Coordinate List (COO) formats, preserving only the values ​​and positions of non-zero elements. The expanded matrix is ​​also normalized to eliminate dimensional differences between modal features, typically using Z-score normalization or Min-Max scaling. For example, a 60-frame, 512×512 pixel ultrasound sequence with three modalities has an original data size of approximately 90MB. After conversion to a sparse matrix representation, the storage requirement can be reduced to 25-35MB, while retaining complete information on non-zero elements, laying the foundation for subsequent principal component analysis.

[0161] Based on the sparse matrix representation, an iterative optimization algorithm is applied to calculate the eigenvalues ​​and eigenvectors of the covariance matrix, and the principal component directions are found under the constraint of sparsity.

[0162] Traditional Principal Component Analysis (PCA) algorithms require calculating the complete covariance matrix and performing eigenvalue decomposition, which is computationally too costly for large-scale medical image data. To address this issue, an iterative algorithm optimized for sparse data is employed to avoid explicitly constructing the complete covariance matrix. Specifically, an improved power method is used to iteratively calculate the principal component directions. First, unit vectors are randomly initialized. V0 As an initial estimate of the principal component direction; then, in each iteration, the matrix-vector product Xv is efficiently calculated by utilizing the properties of sparse matrices. t (where X is the data represented by a sparse matrix), and the result is normalized to obtain the updated vector v. {t+1}The iterative process continues until the vector sequence converges or the maximum number of iterations is reached. To accelerate convergence, a block Krylov subspace method is used, processing multiple vectors in each iteration to form a subspace and extracting the optimal principal component directions. During the computation, sparsity constraints are introduced to ensure that the obtained principal component directions are also sparse. This is particularly important for medical image analysis because sparse principal components are easier to interpret and computationally more efficient. The constraint method uses L1 norm regularization, adding a sparsity penalty term to the objective function: max v T X T Xv - λ||v||1, where λ is an adjustable regularization parameter controlling the sparsity level. An alternating optimization strategy is used to solve this constrained optimization problem: first, the sparsity constraint is fixed, and the principal component directions are optimized; then, the principal component directions are fixed again, and the sparsity pattern is adjusted. Through this iterative optimization process, principal component directions that maximize the data variance explanation rate while satisfying the sparsity requirement are finally obtained, providing an efficient and interpretable foundation for feature extraction from medical images.

[0163] The principal component directions are subjected to soft thresholding. A threshold parameter is set, and coefficients smaller than the threshold parameter are set to zero to maintain sparsity, thus obtaining the principal component coefficients.

[0164] After obtaining the initial principal component directions, their sparse structure needs further optimization to ensure interpretability and computational efficiency. To this end, soft thresholding is employed, a nonlinear operation widely used in sparse signal processing. The core idea of ​​soft thresholding is to retain large coefficients (representing important features) while suppressing small coefficients (potentially representing noise or minor features). Specifically, a soft thresholding function, S, is first applied to each element vi in ​​the principal component direction vector v. λ(vi) = sign(vi)·max(|vi|-λ, 0), where λ is the threshold parameter. This operation sets coefficients with absolute values ​​less than λ to zero, while shrinking coefficients with absolute values ​​greater than λ to reduce their absolute values. The choice of the threshold parameter λ is crucial, and an adaptive strategy is used to determine the optimal threshold: different base thresholds are set for different types of ultrasound examinations (such as heart, abdomen, blood vessels, etc.); the threshold is dynamically adjusted according to the data noise level and the desired sparsity; and the performance under different thresholds is evaluated through cross-validation to select the best balance point. To maintain the orthogonality of the principal components, re-orthogonalization is performed after soft thresholding, usually using a modified Gram-Schmidt process. In addition, a block-based soft thresholding strategy is implemented, applying differentiated threshold parameters to different regions according to the anatomical characteristics of the medical images. For example, more details are preserved in the myocardial region of cardiac ultrasound (using a smaller threshold), while stricter sparsity constraints are applied to the background region (using a larger threshold). Through this intelligent soft thresholding, well-structured principal component coefficients are obtained, which both preserve key diagnostic information and achieve high sparsity. Practice shows that after soft thresholding, 70-85% of the elements in the principal component vector are usually zero, which significantly reduces the subsequent computation and storage requirements while maintaining a variance explanation rate of over 90%.

[0165] The principal component coefficients are updated using the gradient projection method to ensure that the projection results satisfy the sparse constraint conditions. The optimization is iterated until convergence is achieved, and the optimized principal component coefficients are obtained.

[0166] The principal component coefficients after soft thresholding may not be optimal because simple thresholding operations can destroy some important properties of the principal components. To further optimize the principal component coefficients, an iterative optimization using the gradient projection method is employed. This method combines the efficiency of gradient descent with the constraint-satisfying ability of projection operations, making it particularly suitable for solving constrained optimization problems. In the specific implementation, the system first defines the objective function J(v) = -v T X T Xv, which means maximizing the negative of the projection variance (transformed into a minimization problem). In each iteration, the objective function is calculated with respect to the current principal component coefficients v. t The gradient ∇J(v) t ) = -2X T X v t Then update the coefficients along the gradient direction: v'{t+1} = v t - α t ∇J(v t ), where α tThe adaptive step size is determined via line search or the Armijo criterion. The updated coefficients v'{t+1} may not satisfy the constraints, therefore the system performs a projection operation, projecting them onto the constraint set: v {t+1} = P C (v' {t+1} ), where P C Let C represent the projection operator, and C be the constraint set (usually an L1 norm sphere or sparsity constraints). To improve computational efficiency, an approximate projection method is employed, such as the Iterative Soft Thresholding Algorithm (ISTA) or its accelerated version, FISTA. The iterative process continues until the convergence condition is met: the change in principal component coefficients between two consecutive iterations is less than a preset threshold ε, or the maximum number of iterations is reached. To handle multiple principal components, a progressive strategy is adopted: after calculating the first principal component, its contribution is subtracted from the data, and then the next principal component is calculated in the residual space, ensuring the orthogonality between the principal components. A parallel computing mechanism is also implemented, utilizing a multi-core processor to simultaneously optimize multiple principal components, significantly improving computational efficiency. Through this refined iterative optimization process, optimized principal component coefficients that satisfy both sparsity constraints and maximize variance explanation rate are finally obtained, providing a solid foundation for efficient compression and feature extraction of medical images.

[0167] Based on the optimized principal component coefficients, the compressed feature data is reconstructed, and the compression ratio and feature retention evaluation index are calculated to obtain the compressed feature data.

[0168] After obtaining the optimized principal component coefficients, the final data reconstruction and evaluation stage begins. First, the original sparse matrix X is projected onto a low-dimensional subspace composed of the first k principal components, resulting in the eigenvalue matrix Y = XV, where V is a matrix containing k principal component vectors. These eigenvalues ​​represent the coordinates of the original data in the principal component space, capturing the main variation patterns of the data. The number of principal components k to be retained is adaptively determined based on application requirements and data characteristics, typically using the cumulative variance explained rate criterion (retaining principal components that explain 90-95% of the total variance) or the eigenvalue truncation criterion (retaining principal components with eigenvalues ​​greater than a preset threshold). Different retention strategies are set for different types of ultrasound examinations: for examinations requiring detailed structural analysis (such as heart valve assessment), more principal components are retained; for examinations primarily focusing on gross morphology (such as routine abdominal screening), the number of principal components can be appropriately reduced. After projection, the storage requirements of the compressed data, including the eigenvalue matrix Y and the principal component matrix V, are calculated. Compared to the original data, the compression ratio CR is calculated as: Original data size / Compressed data size. For typical ultrasound sequences, compression ratios are usually between 4:1 and 10:1, depending on data characteristics and the number of principal components retained. To evaluate compression quality, several feature retention metrics are calculated: reconstruction error (mean squared error between the original and reconstructed data), typically required to be less than a preset threshold; structural similarity index (SSIM), measuring the degree of preservation of image structural information, typically required to be greater than 0.9; and fidelity of diagnostic key regions, assessed by comparison with key regions annotated by medical experts. Visual quality assessment is also performed, allowing medical staff to evaluate clinical usability by displaying the original and reconstructed images side-by-side. All these evaluation metrics are combined to form a quality score. When the score exceeds a preset threshold, compression is confirmed as effective, and the final compressed feature data is generated, including feature coefficients, principal component vectors, and necessary metadata (such as compression parameters and evaluation metrics). This compressed feature data not only significantly reduces storage requirements but also provides high-quality input after dimensionality reduction for subsequent AI analysis, significantly improving system performance and user experience in offline environments.

[0169] Example 7

[0170] Based on Example 3, the connected component analysis of the temporal correlation graph is performed to identify missing frames, corrupted frames, and temporally anomalous regions in the image sequence, and the number and topological characteristics of connected components are counted to obtain the connected component analysis results, including:

[0171] The time-series correlation graph is traversed using a depth-first search algorithm to identify the set of interconnected nodes and obtain the connected components and the number of connected components.

[0172] When assessing the integrity of ultrasound image sequences, it is first necessary to determine the connectivity between frames in the sequence. To this end, a graph traversal algorithm based on Depth-First Search (DFS) was implemented to systematically explore the structure of the temporally correlated graph. Specifically, the system first sets an edge weight threshold τ (typically 0.4-0.6, dynamically adjusted according to different examination types), temporarily removing edges with weights below τ from the graph to filter out weakly correlated connections. Then, an unvisited starting node is selected, and the standard DFS process is executed: the current node is marked as "visited," and all its unvisited adjacent nodes are recursively visited until no further expansion is possible. After one DFS traversal, all visited nodes form a connected component. A new starting point is selected from the remaining unvisited nodes, and the DFS process is repeated until all nodes in the graph have been visited. To improve algorithm efficiency, an iterative implementation is used instead of a recursive implementation to avoid stack overflow issues that may occur in large-scale sequences; simultaneously, an adjacency list is used to store the graph structure, optimizing memory usage and access speed. For each identified connected component, the system records its node set, number of nodes, temporal span (interval between the earliest and latest frames), and spatial coverage. It also calculates the density of the connected component, the ratio of the actual number of edges to the theoretical maximum number of edges, as an indicator of the tightness of the connection. Finally, the system obtains the total number of connected components and detailed information for each component. Ideally, a complete image sequence should form a single connected component; the presence of multiple connected components indicates a break in the sequence. For example, in a 120-frame echocardiogram, if three connected components are identified, containing 45, 38, and 37 frames respectively, it indicates a break in the sequence at two locations, requiring further analysis to determine the cause and location of the break.

[0173] For each connected component, topological characteristic parameters, including connectivity, diameter, clustering coefficient, and centrality index, are calculated to evaluate the integrity of the image sequence and obtain integrity results.

[0174] After identifying the connected components, a series of graph theory metrics are further calculated to analyze the topological properties of each component in depth. First, connectivity is calculated, which is the minimum number of edges that need to be removed to reach another node from one node. This reflects the robustness of the sequence—high connectivity indicates that the overall sequence remains connected even if some inter-frame relationships are disrupted. Second, diameter is calculated, which is the maximum value of the shortest path between any two points in the graph. A smaller diameter indicates that the "distance" between any two frames in the sequence is closer, and the temporal continuity is better. Third, the average clustering coefficient is calculated, which measures the degree of interconnection between the neighbors of nodes in the graph. A high clustering coefficient indicates the existence of a tight local structure in the sequence, such as the systolic and diastolic frame sets in a cardiac cycle. Fourth, various centrality metrics are calculated: degree centrality identifies the most connected key frames; betweenness centrality identifies "bridging frames" connecting different subsequences; and eigenvector centrality identifies frames closely connected to other important frames. Based on these topological parameters, a completeness scoring model is constructed to comprehensively evaluate the completeness of a sequence. This model uses a weighted summation method: Completeness = W1 ×Connectivity+ W2 ×(1 / Diameter) + W3 ×ClusteringCoeff + W4 ×CentralityPattern, where weighting coefficients are optimized based on different examination types. For example, for cardiac ultrasound, periodic structures are more important, so the clustering coefficient has a higher weight; while for static organ scans, connectivity has a higher weight. The scoring results are standardized to 0-100 points and divided into four levels: "Excellent (90-100)," "Good (75-90)," "Average (60-75)," and "Poor (0-60)." Connected components with scores below the threshold (usually 75 points) are marked as "Needs Improvement" and a specific problem description is generated, such as "Insufficient Connectivity" or "Local Structure Missing," etc.

[0175] By analyzing the temporal span and spatial distribution characteristics of the connected components, the location of sequence breakage caused by equipment failure, transmission interruption, or buffer overflow can be identified.

[0176] After confirming the existence of a break in the sequence, further analysis of the specific location and possible causes of the break is needed to provide a basis for subsequent repair. First, the time label distribution of each connected component is analyzed to construct a time coverage map and identify "blank areas" on the time axis, i.e., the time periods when frames are missing. For each break point, the boundary features of the connected components before and after the break are calculated: time interval (break duration), spatial displacement (difference in the anatomical structure position of frames before and after the break), and feature similarity (the degree of similarity in image features between frames before and after the break). Based on these features, a decision tree classifier is applied to identify the cause of the break: large time intervals (>500ms) and high feature similarity usually indicate a transmission interruption; small time intervals but large spatial displacements and low feature similarity usually indicate sudden device movement; breaks occurring near the buffer size and involving multiple consecutive missing frames may be caused by buffer overflow. Quality metrics of frames near the break point, such as signal-to-noise ratio, contrast, and sharpness, are also analyzed to detect patterns of gradual quality degradation followed by sudden recovery, which is often a characteristic of temporary device malfunctions. For periodic sequences (such as echocardiography), it is also detected whether the break occurs at a specific phase of the physiological cycle. For example, if multiple cardiac cycles show breaks during similar cardiac contraction phases, it may indicate a systemic problem with the device when capturing rapid motion. All these analytical results are integrated into a structured break description, including break location (start and end frame indices), duration, possible causes (ranked by probability), and severity rating. For example, it might report, "A moderate-severity sequence break was detected between frames 45 and 48, lasting approximately 200 ms. The most likely cause is transient ultrasound probe movement (85% confidence level), and repair via periodic interpolation is recommended." This detailed break analysis provides precise guidance for subsequent sequence repair, enabling the selection of the most appropriate repair strategy.

[0177] The rationality of the connected components is verified based on the physiological periodicity pattern. For periodic images, the connected patterns corresponding to the physiological cycle are detected, and physiological verification results are obtained.

[0178] Medical ultrasound images, especially cardiac and vascular ultrasound, typically exhibit distinct physiological periodicity. This characteristic is utilized to further validate the physiological rationale of connected components. First, the system applies periodicity detection algorithms, including autocorrelation and Fast Fourier Transform (FFT), to each connected component to identify potential periodic patterns. For cardiac ultrasound, the motion trajectories of key anatomical landmarks (such as ventricular diameter and valve opening / closing position) are extracted, and the period length and phase information of these trajectories are calculated. Ideally, a complete cardiac cycle should appear as a complete closed loop of characteristic trajectories; the number of complete cycles and partial cycle segments contained in each connected component are calculated. For partial cycles, their physiological stage is further identified, such as systole, early diastole, mid-diastole, or late diastole. The consistency between cycles is also analyzed, calculating the similarity and coefficient of variation between consecutive cycles; abnormally high variation usually indicates sequence problems. For ultrasound of non-periodic organs (such as routine abdominal ultrasound), validation is performed based on the continuity of anatomical structures and the smoothness of the scanning path. A scanning trajectory model is constructed to analyze the continuity and coverage patterns of probe movement, identifying abnormal skipping or repetitive regions. Based on these analyses, a physiological validation score is generated, reflecting the degree to which connected components conform to the expected physiological pattern. For periodic sequences, the score considers the proportion of complete cycles, cycle consistency, and the presence of key physiological events (such as peak cardiac contraction); for non-periodic sequences, the score considers the integrity and smoothness of the scan path. The system standardizes the scores to 0-100 points and sets a threshold (usually 80 points) to determine whether connected components are physiologically reasonable. Connected components with scores below the threshold are marked as "physiologically unreasonable" and provided with a specific problem description, such as "missing complete systolic phase" or "discontinuous scan path." This physiologically based validation significantly improves the accuracy and clinical relevance of the system's assessment.

[0179] Based on the number of connected components, the integrity result, the sequence break location, and the physiological verification result, a connected component analysis result is constructed.

[0180] Integrating the aforementioned analyses, a comprehensive connected component analysis is constructed, providing a systematic basis for sequence quality assessment and subsequent processing. The analysis results are organized hierarchically, comprising two levels: global assessment and component-level details. At the global assessment level, the system provides an overview of the overall sequence quality: the total number of connected components and their distribution characteristics, such as size distribution, temporal coverage, and spatial distribution patterns; a total sequence integrity score, comprehensively considering the integrity scores and coverage of all connected components; the total number of breaks and their distribution patterns, such as whether they are concentrated in specific time periods or physiological stages; and a total physiological rationality score, reflecting the degree of conformity between the overall sequence and the expected physiological pattern. At the component-level details level, a detailed analysis is provided for each connected component: basic attributes, including the number of nodes, time span, and start and end frame indices; topological characteristic parameters, including the aforementioned connectivity, diameter, clustering coefficient, and centrality index; temporal characteristics, including frame rate stability, temporal coverage uniformity, and potential periodic patterns; spatial characteristics, including anatomical structure coverage and scanning path characteristics; physiological characteristics, including the number of included physiological cycles, cycle integrity, and coverage of key physiological events; and quality indicators, including average signal-to-noise ratio, contrast, and sharpness. The system also generates analyses of relationships between connected components, including temporal relationships, spatial overlap, and content similarity. This information is crucial for understanding the nature of sequence breaks and planning repair strategies. All these analysis results are stored in structured JSON format for easy parsing and utilization by subsequent processing modules. Simultaneously, visual reports are generated, including intuitive representations such as connected component distribution maps, temporal coverage heatmaps, and key parameter radar charts, helping users quickly understand the sequence status. This comprehensive and structured analysis provides a solid foundation for subsequent sequence repair and quality control.

[0181] The connected component analysis results are compared with a preset integrity threshold. When the connectivity of any connected component is lower than the preset threshold, it is marked as an abnormal sequence that needs to be reconstructed, and a sequence quality assessment report is generated.

[0182] After comprehensive analysis, the final evaluation and decision-making stage begins, determining whether the sequence needs reconstruction and the specific measures to be taken. First, key indicators for each connected component are compared with preset thresholds, which are differentiated based on different examination types and clinical requirements. Core evaluation indicators include: connectivity threshold τc, typically set at 50-60% of the average node degree; integrity score threshold τi, typically set at 75-80 points; physiological rationality score threshold τp, typically set at 80-85 points; and minimum effective node number threshold τn, set according to the examination type, such as cardiac ultrasound, which typically requires at least one complete cardiac cycle frame. When the connectivity of any connected component is lower than τc, or the integrity score is lower than τi, or the physiological rationality score is lower than τp, or the number of nodes is less than τn, it is marked as an "abnormal sequence requiring reconstruction." For the labeled connected components, the system further analyzes the severity and nature of the problems, generating tiered processing suggestions: for minor problems (such as small breaks or connectivity close to the threshold), basic interpolation repair is recommended; for moderate problems (such as obvious breaks but basic physiological characteristics are preserved), periodic reconstruction is recommended; for severe problems (such as large-scale missing data or multiple breaks), re-acquisition or labeling as "for reference only, not for diagnosis" is suggested. These assessment results and suggestions are integrated into a structured sequence quality assessment report, containing the following main parts: basic sequence information, such as patient ID, examination type, acquisition time, and equipment information; a quality assessment summary, including overall score, number of problems, and severity distribution; a detailed list of problems, including the location, type, severity, and recommended processing method for each problem; a visualization, including a problem area marker map and a quality distribution heatmap; and a summary of processing suggestions, including whether reconstruction is needed, recommended reconstruction methods, and expected outcome assessment. This report serves as input to the system's internal reconstruction module, guiding the automated repair process, and can also be provided to physicians for review, helping them understand the data quality status and the basis for system processing decisions, ensuring that clinical decisions are based on a thorough understanding of data quality.

[0183] Example 8

[0184] Based on Example 4, the quantitative measurement and analysis based on the lesion area detection results, calculating geometric parameters such as lesion area, perimeter, major axis, and minor axis, and generating structured measurement data includes:

[0185] The detection results of the lesion area are processed by boundary extraction, and the lesion contour is identified by edge detection algorithm and morphological filtering is performed to remove noise.

[0186] After obtaining preliminary detection results of the lesion region, accurate boundary extraction is required, which is the foundation for subsequent quantitative measurements. A multi-stage boundary extraction strategy is adopted to ensure accurate lesion contours in noisy and blurred medical images such as ultrasound images. First, the detected lesion region is preprocessed, including adaptive histogram equalization to enhance contrast, anisotropic diffusion filtering to suppress noise while preserving edge information, and wavelet transform-based multi-scale enhancement. These preprocessing steps significantly improve the visibility of lesion boundaries. Second, an integrated strategy of multiple edge detection algorithms is applied: the Canny operator provides basic edge detection, and its dual-threshold design is particularly suitable for handling weak boundaries in ultrasound images; the Sobel and Prewitt operators provide directional gradient information, which helps to identify edges in different directions; and the LoG (Laplacian of Gaussian) operator provides second-order derivative information to enhance edge localization accuracy. The system integrates the results of these algorithms through a weighted voting mechanism to fully utilize the advantages of each algorithm. For the speckle noise unique to ultrasound images, specialized edge enhancement algorithms, such as phase consistency enhancement and local phase enhancement, are implemented. These methods have strong robustness to noise. After obtaining the initial edges, a series of morphological operations are applied for refinement: morphological opening and closing operations remove small noise points and fill small holes; edge connection algorithms repair broken edges based on directional consistency and distance constraints; edge thinning algorithms (such as the Zhang-Suen thinning algorithm) transform coarse edges into precise contours with a single pixel width. To handle the common blurred boundaries in ultrasound images, an active contour model based on the level set method is implemented, which gradually conforms the contour to the real boundary through the principle of energy minimization. A boundary optimization module based on deep learning is also integrated, using pre-trained U-Net or DeepLab models to further refine the boundaries. Finally, curve smoothing algorithms (such as B-spline interpolation or Gaussian smoothing) are applied to smooth the extracted contours, eliminating jagged effects and local irregularities while preserving the overall morphological features of the contours. Through this multi-stage, multi-algorithm fusion boundary extraction process, accurate and smooth lesion contours can be obtained from complex ultrasound images, laying a solid foundation for subsequent quantitative measurements.

[0187] The area parameters are calculated based on the lesion contour, and the number of pixels in the lesion area is obtained through pixel statistics. The actual area value is then converted by combining the pixel spacing.

[0188] After obtaining a precise lesion outline, the area calculation stage begins, which is a fundamental parameter for assessing lesion size. Several complementary area calculation methods were implemented to ensure the accuracy and reliability of the measurements. First, a pixel counting method was used to calculate the basic area: the area inside the lesion outline was marked as the foreground, and the total number of foreground pixels was counted; for pixels on the outline boundary, sub-pixel precision processing was used, assigning weights based on the proportion of pixels covered by the outline to improve the accuracy of the area calculation. Special attention was paid to the unique characteristics of ultrasound images, and targeted optimizations were implemented: for acoustic shadowing areas (usually appearing behind calcifications or high-density tissues), an intelligent compensation algorithm was used to estimate the true boundary of the obscured area based on surrounding tissue features and anatomical knowledge; for echo-enhanced areas (usually appearing behind cystic structures), specific boundary correction rules were applied to avoid including enhancement artifacts in the area calculation. After obtaining the pixel count, physical unit conversion was performed, converting the pixel count into actual area values ​​(usually in square millimeters). This process requires accurate pixel spacing information, which was extracted from the DICOM header information or calibrated using a scale in the image. For different ultrasound probes and scanning depths, the conversion parameters were dynamically adjusted to ensure measurement consistency. To improve the reliability of area calculation, a multi-method verification mechanism is implemented: in addition to direct pixel counting, Green's theorem (based on contour integral calculation) and polygon approximation (approximating the contour as a polygon and calculating the area using geometric formulas) are also used to calculate the area. The results of different methods are compared, and a warning is triggered and manual review is recommended when the difference exceeds a threshold (usually 3%). The scanning plane factor of ultrasound imaging is also considered; for scanning planes not perpendicular to the tissue surface, an angle correction option is provided to geometrically correct the area based on the scanning angle. For 3D ultrasound data, a volume calculation function is implemented: the total volume is calculated by measuring the area on multiple parallel cross-sections and then applying numerical integration methods (such as the trapezoidal rule or Simpson's rule). A detailed area calculation report is generated, including the original pixel count, conversion parameters, correction factors, and final area value, ensuring the transparency and traceability of the measurement process.

[0189] The perimeter of the lesion contour is measured, the perimeter of the lesion contour is calculated using a contour tracking algorithm, and a curve smoothing algorithm is applied to eliminate the jagged effect.

[0190] Perimeter is a crucial parameter for assessing lesion morphology, especially in differentiating between benign and malignant lesions, where the complexity of the boundary often has significant diagnostic value. A high-precision perimeter measurement process was implemented, fully considering the discrete characteristics of contour representation in digital images. First, a contour tracking algorithm (such as Moore's neighborhood tracking algorithm or Pavlidis contour tracking algorithm) is applied to track the lesion boundary pixel by pixel, generating an ordered sequence of boundary points. During tracking, two modes are distinguished: four-connected and eight-connected. The eight-connected mode is typically used to obtain a smoother contour representation. For locally unclear boundaries commonly seen in ultrasound images, an adaptive tracking strategy is implemented: high-precision tracking is used in areas with clear boundaries, while inference methods based on local features and global shape constraints are used to complete the boundaries in blurred areas. After obtaining the boundary point sequence, the contour is encoded using chain code representation (Freeman chain code). This compact representation facilitates subsequent analysis and processing. Based on the chain code representation, the initial perimeter is calculated: horizontal and vertical movements (chain codes 0, 2, 4, 6) are counted as 1 unit length, and diagonal movements (chain codes 1, 3, 5, 7) are counted as... The total length is calculated by multiplying the pixel spacing to convert it to physical units (millimeters). However, this direct calculation method leads to a "jagged effect," causing the perimeter measurement to be too large. To address this issue, various curve smoothing and perimeter correction techniques are applied: Kulpa correction, which corrects the chain code perimeter based on statistical principles, typically setting the correction coefficient to 0.948; polynomial fitting smoothing, which uses a piecewise cubic polynomial (such as a B-spline curve) to fit the original contour and then calculates the length of the smoothed curve; and Fourier descriptor smoothing, which transforms the contour to the frequency domain, filters out high-frequency components, and then transforms it back to the spatial domain to obtain a smooth contour. The system also implements curvature-based adaptive smoothing, preserving more details in high-curvature regions (such as sharp corners) and performing stronger smoothing in low-curvature regions. This method is particularly suitable for preserving the characteristic morphology of lesion boundaries. For boundary irregularities of particular clinical concern, the perimeter ratio (the ratio of the actual perimeter to the circumference of a circle of equal area) is calculated. This dimensionless parameter is an important indicator for assessing boundary complexity. Generate a perimeter measurement report, including the original perimeter value, the corrected perimeter value, the correction method used, and the perimeter ratio, providing a comprehensive reference for clinical assessment.

[0191] The major and minor axes of the lesion contour are calculated using the minimum bounding rectangle method to determine the main orientation and geometric parameters of the lesion contour.

[0192] The major and minor axes are fundamental parameters describing the size and shape of lesions and play a crucial role in various ultrasound diagnostic standards. A precise measurement method based on the minimum bounding rectangle was implemented to ensure consistent and reliable measurement results for lesions of various morphologies. First, the orientation of the lesion contour is calculated, which determines the orientation of the major axis. Principal component analysis (PCA) is used: the contour points are treated as a two-dimensional point set, and its covariance matrix is ​​calculated. The direction of the first eigenvector is the principal orientation of the lesion. This method can find the orientation with the greatest data variability and is suitable for most non-circular lesions. For lesions that are close to circular, the stability of the principal orientation is tested. When two eigenvalues ​​are close (ratio > 0.9), it is marked as "approximately circular," in which case the orientation of the major axis has some arbitrariness. After determining the principal orientation, the minimum bounding rectangle is calculated: the contour point set is rotated until the principal orientation is aligned with the coordinate axes, the maximum and minimum coordinate values ​​of the rotated point set are found, the bounding rectangle is constructed, and then the rectangle is rotated back to its original orientation. This method ensures that the sides of the bounding rectangle are aligned with the principal orientation of the lesion, and the long and short sides of the rectangle correspond to the major and minor axes of the lesion, respectively. To improve measurement accuracy, the system implements several optimization techniques: for irregular shapes, multiple candidate directions are tried (e.g., based on moment of inertia, based on principal axis of convex hull), and the direction that produces the smallest bounding rectangle is selected; for shapes with obvious depressions, the influence of the depressions on the principal direction is considered, and morphological closing operations are performed to fill small depressions before calculating the principal direction if necessary. After obtaining the smallest bounding rectangle, its long and short sides are directly measured and used as the major and minor diameters of the lesion, respectively. The major-to-minor diameter ratio (major diameter / minor diameter) is also calculated, which is an important indicator for assessing the elongation of the lesion shape and has reference value in distinguishing between benign and malignant lesions. In addition to rectangle-based measurements, complementary measurement methods are also implemented: maximum inscribed circle diameter, which finds the diameter of the largest circle that can be completely contained within the lesion; maximum Feret diameter, which is the maximum distance between any two points on the lesion; and average diameter, which calculates the average diameter in multiple directions passing through the centroid of the lesion. The system generates a detailed diameter measurement report, including major diameter, minor diameter, major-to-minor diameter ratio, principal direction angle (relative to the horizontal direction of the image), and various complementary diameter measurements, providing a comprehensive reference for clinical evaluation. These geometric parameters are not only used for size assessment, but also form the basis for subsequent shape analysis.

[0193] Based on the geometric parameters, a shape descriptor is calculated, including morphological indices such as roundness, slenderness, and irregularity, to construct the structured measurement data containing location, size, shape, and texture features.

[0194] After completing basic geometric measurements, the advanced morphological analysis stage begins, calculating a series of shape descriptors. These indicators are crucial for differentiating between benign and malignant lesions and determining the nature of lesions. First, circularity is calculated: Circularity = 4π × area / perimeter², ranging from 0 to 1. A value closer to 1 indicates a shape closer to a perfect circle. Circularity is the most fundamental shape indicator and is widely used in various ultrasound diagnostic standards. For example, high circularity in breast masses usually suggests benignity, while low circularity increases the likelihood of malignancy. Second, elongation is calculated: Elongation = 1 - minor axis / major axis, ranging from 0 to 1. A larger value indicates a more elongated shape. Elongation reflects the extent of the lesion's extension in the principal direction. Some types of benign lesions (such as fibroadenomas) are typically elliptical and have a moderate elongation. Third, Irregularity is calculated: Irregularity = actual perimeter / convexity perimeter, typically ranging from 1 to greater than 1. A larger value indicates more irregular boundaries. Irregularity is an important indicator for assessing boundary complexity, and high irregularity is often associated with malignant lesions. A series of high-level shape descriptors are also calculated: Convexity = Area of ​​the convex hull / Actual area, reflecting the degree of convexity or concavity of the shape; Compactness = √(4 × area / π) / Major axis, reflecting the degree of concentration of the shape; Rectangularity = Area / Area of ​​the smallest bounding rectangle, reflecting the similarity between the shape and a rectangle; Fractal Dimension, calculated using box counting, reflects the complexity and self-similarity of the boundary. For ultrasound-specific morphological features, specialized analyses are implemented: Posterior echo feature analysis to assess acoustic attenuation or enhancement patterns behind the lesion; Edge ambiguity analysis to quantify the clarity of the lesion boundary and the width of the transition zone; and Calcification feature analysis to detect and describe the distribution pattern of hyperechoic points within the lesion. All measurement and analysis results are integrated into a structured measurement data package containing four main parts: location information, including the lesion's coordinates in the image, anatomical location, and relative position to key anatomical landmarks; size information, including area, perimeter, various diameter measurements, and volume estimates (where applicable); shape information, including all shape descriptors and morphological feature analysis; and texture information, including quantitative descriptions of echo patterns, internal homogeneity, and specific echo characteristics. This comprehensive structured data not only supports the generation of current diagnostic reports but also facilitates long-term follow-up comparisons and big data analysis, providing rich and objective evidence for clinical decision-making.

[0195] Example 9

[0196] Based on Example 5, the step of applying the maximum weighted independent set algorithm to optimize the synchronization strategy based on the sparse graph structure, and solving for the data subset with the largest weight that can be synchronized in parallel using a distributed approximation algorithm, includes:

[0197] The priority weights of the data to be synchronized are calculated using a multi-factor weighted calculation method, which includes data priority factor, time urgency factor and data size influence factor.

[0198] After network recovery, a large amount of accumulated offline data needed to be synchronized to the central server. To ensure the priority transmission of the most important medical data, a refined multi-factor weighted priority calculation mechanism was implemented. First, a basic priority factor (P) was assigned based on data type and clinical importance: emergency examination data was given the highest weight (P=10) to ensure priority synchronization of critically ill patient information; routine medical data was given a medium weight (P=5) to ensure the continuity of daily medical activities; and physical examination and research data were given a lower weight (P=1) to allow for appropriate delays in synchronization when resources are limited. Second, the system calculated a time urgency factor (T) using a non-linear time decay function: T = 1 / (1+α·(t)). current -t create )), where t current t represents the current time. create For data creation time, α is the decay coefficient. This design ensures that newly generated data receives higher priority, while non-urgent data that has not been synchronized for a long time gradually decreases in priority. Third, considering the data size impact factor (S), a piecewise function is used: for small data (<1MB, such as text reports, measurement results), S=1.2; for medium data (1-10MB, such as single-frame images), S=1.0; for large data (>10MB, such as ultrasound sequences), S=0.8. This design encourages the system to prioritize the transmission of small, critical data, improving initial synchronization efficiency. Finally, the comprehensive weight is calculated through a weighted summation: W = α×P + β×T + γ×S, where α, β, and γ are configurable weight coefficients, with default settings of α=0.5, β=0.3, and γ=0.2. An adaptive weight adjustment mechanism is also implemented, dynamically adjusting the weight coefficients based on network conditions and server load: increasing the γ value under bandwidth constraints to prioritize the transmission of small files; increasing the β value during peak server load periods to prioritize the processing of the latest data. Through this multi-dimensional weight calculation, the system can make intelligent data synchronization decisions in complex and ever-changing medical environments.

[0199] The graph is constructed based on data dependencies, including the association dependency between reports and images, the mapping dependency between patient IDs and examination data, and the conflict constraints of equipment resources.

[0200] After calculating node weights, it is necessary to clarify the various relational constraints between data, as these constraints directly affect the feasibility of the synchronization strategy. A complex edge connection structure is constructed to comprehensively express the multidimensional relationships between medical data. First, logical dependency edges are established to represent the sequential constraints between data. These edges include: the association dependency between reports and images, ensuring that image data referenced in diagnostic reports is synchronized before the report itself, avoiding the reference of non-existent images in the report; the mapping dependency between patient information and examination data, ensuring that the patient's master index record is synchronized before the patient's examination data, maintaining database reference integrity; the inclusion dependency between examination records and image sequences, ensuring that examination metadata is synchronized before the specific image sequence generated by the examination; and the derivation dependency between measurement results and measurement objects, ensuring that raw data is synchronized before measurement results generated based on that data. Second, resource conflict edges are established to represent data pairs that are not suitable for parallel processing. These edges include: conflicts between large DICOM files generated by the same ultrasound equipment, avoiding I / O bottlenecks caused by parallel reading; conflicts between multiple status update records of the same patient, preventing data overwrite risks caused by parallel processing; and conflicts between multiple records sharing the same large attachment, avoiding duplicate transmission of the same content. Third, a priority preemption edge is established, indicating that high-priority data can preempt the transmission resources of low-priority data. For example, newly generated emergency ultrasound examinations can preempt historical physical examination data that is being synchronized. The system assigns different weights and attributes to these edge relationships. For example, dependency edges are marked as "mandatory constraints," and violations will lead to data inconsistency; while resource conflict edges are marked as "performance constraints," and violations will lead to decreased efficiency but not affect accuracy. Through this refined edge connection structure, the system constructs a graph model that comprehensively reflects the complex relationship network of medical data, providing accurate constraints for subsequent optimization algorithms.

[0201] A distributed algorithm is applied for multi-round iterative computation. In each round, each node updates its own state based on local information and coordinates global optimization through a message passing mechanism.

[0202] After constructing the weighted graph structure, the maximum weighted independent set problem needs to be solved on the graph to determine the optimal parallel synchronization strategy. Considering the distributed nature of medical data (potentially scattered across multiple ultrasound workstations) and the complexity of the problem, a distributed approximation algorithm based on the Congest model is adopted. This algorithm is particularly suitable for execution in environments with limited communication bandwidth, with each message size being only O(log n) bits, where n is the number of nodes. The algorithm execution process consists of multiple iterations, each round containing four key steps: information collection, decision value calculation, state update, and result propagation. In the information collection phase, each node broadcasts its current state (weight, whether it has been selected or excluded) and node ID to its one-hop neighbors. In complex network environments, a reliable message passing mechanism is implemented, including message acknowledgment, timeout retransmission, and sequence number management, ensuring reliable information transmission in unstable networks. During the decision value calculation phase, each node calculates its own decision value D(v) = w(v) / f(d(v)) based on the collected local information, where w(v) is the node weight, d(v) is the number of unprocessed neighbors, and f is a monotonically increasing function (usually f(d) = d + 1 or f(d) = ... This design tends to select nodes with "high weight and few constraints," balancing greedy selection and global optimization. During the state update phase, the system employs a probabilistic decision-making mechanism: the node with the highest decision value is selected with a high probability (typically >0.8) to join the independent set; all neighbors of the selected node are marked as "excluded"; the remaining nodes remain in a "pending" state. This probabilistic design helps avoid local optima and improves the algorithm's global optimization capability. During the result propagation phase, nodes whose states have changed broadcast their new states to their neighbors, triggering the next iteration. To improve algorithm efficiency, several optimizations are implemented: adaptive iterative control, dynamically adjusting the number of iterations based on network conditions and graph structure complexity; local preprocessing, pre-determining the optimal order for explicit linear dependency chains; and parallel execution optimization, fully utilizing multi-core processing capabilities to process multiple subgraphs simultaneously. Through this distributed iterative computation, the maximum weighted independent set problem can be solved efficiently in complex constrained networks, providing an optimization strategy for data synchronization.

[0203] In each iteration, the candidate set of nodes is calculated, and the data item with the largest weight that does not conflict with the selected node is added to the independent set, until it can no longer be expanded.

[0204] The core of distributed algorithms lies in how to efficiently select the optimal node to join the independent set in each iteration. A sophisticated candidate set calculation and node selection mechanism is implemented to ensure the algorithm's efficiency and result quality. At the beginning of each iteration, the current candidate node set is determined, consisting of all nodes in the "pending" state that are not adjacent to any selected node. To improve computational efficiency, an incremental update strategy is adopted instead of recalculating the entire candidate set each time: when a node is selected, only its neighbors need to be removed from the candidate set; after an iteration ends, the newly added nodes are checked, and the candidate eligibility of their two-hop neighbors is updated. For each node in the candidate set, the system calculates its effective decision value, considering node weights, neighbor constraints, and global balance: EV(v) = w(v) × (1 - Σ(u∈N(v)) p(u)) / (d(v) + 1), where N(v) is the set of neighbors of v, and p(u) is the probability estimate of neighbor u being selected. This design considers not only the node's own weight but also the probability of its neighbors being selected, avoiding suboptimal solutions that may result from greedy selection. In the node selection phase, instead of simply selecting the node with the highest decision value, a randomized selection strategy is adopted: candidate nodes are arranged in descending order of decision value, and the probability of node v being selected is proportional to exp(λ·EV(v)), where λ is a temperature parameter that controls the degree of randomness. Initially, the value of λ is small, introducing higher randomness to explore the solution space; as iterations proceed, the value of λ gradually increases, and the algorithm's behavior tends towards greedy selection, accelerating convergence. To handle large-scale datasets, a partitioning strategy is implemented: the graph is divided into multiple overlapping subgraphs, node selection is performed in parallel on each subgraph, and then the results are merged and boundary conflicts are resolved. An adaptive termination condition is also implemented: the algorithm terminates when the candidate set size changes less than a threshold (usually 1%) over multiple consecutive iterations (typically 3-5 rounds), or when the preset maximum number of iterations (usually log(n) rounds, where n is the number of nodes) is reached. This sophisticated candidate set calculation and node selection mechanism can find near-optimal weighted independent sets under complex constraints, providing a high-quality strategy for medical data synchronization.

[0205] The feasibility of the calculated maximum weighted independent set is verified to ensure that the selected data items meet the parallel synchronization conditions and do not violate the dependency constraints, and the optimal synchronization strategy scheme is generated.

[0206] After calculating the maximum weighted independent set, the algorithm enters the crucial verification and optimization phase to ensure the feasibility and effectiveness of the synchronization strategy in a real-world environment. First, a comprehensive dependency verification is performed to check for violations of mandatory dependency constraints. Specifically, a dependency graph of the selected nodes is constructed, and a topology sorting algorithm is executed to verify for circular dependencies. If a dependency violation is found, a minimum correction strategy is adopted: calculating which nodes to remove minimizes weight loss while resolving all dependency conflicts, typically implemented using dynamic programming. Second, resource utilization balance is evaluated to check whether the synchronization strategy will overload certain resources (such as I / O channels of specific devices, network bandwidth, or server processing power). The expected load distribution for each resource is calculated, and when a potential bottleneck is detected, the load is balanced by adjusting the execution time windows of some nodes without removing them from the independent set. Third, transmission time estimation and batch partitioning are performed. Based on data size, network bandwidth, and server processing power, the transmission time for each data item is estimated. The system divides the nodes in the independent set into multiple transmission batches, controlling the total transmission time within each batch within a preset threshold (usually 30-60 seconds) to avoid prolonged resource occupation. Batch partitioning considers node weights and dependencies, ensuring that high-weight nodes are processed first while satisfying dependency constraints. Fourth, a detailed execution plan is generated, assigning specific execution parameters to each data item: start time window, retry strategy, timeout settings, and rollback scheme. For critical data, the system configures a more aggressive retry strategy and a longer timeout; for non-critical data, a more conservative setting is adopted, quickly releasing resources in case of failure. Finally, the verified and optimized results are integrated into a complete synchronization strategy scheme, comprising three core parts: a batch partitioning table, which details the data items included in each batch and their priorities; an execution parameter table, which specifies detailed execution control parameters for each data item; and a monitoring metric table, which defines key metrics that need to be monitored in real time and their thresholds for dynamically adjusting the execution plan. This comprehensive verification and optimization ensures that the synchronization strategy generated by the system is not only theoretically optimal but also executes efficiently and reliably in complex real-world environments.

[0207] Example 10

[0208] Based on Example 6, the compressed feature data is reconstructed using the optimized principal component coefficients, and the compression ratio and feature retention evaluation index are calculated to obtain the compressed feature data, including:

[0209] Image data is reconstructed based on the principal component coefficients, and compressed features are mapped back to the original image space through matrix multiplication.

[0210] After completing principal component analysis and coefficient optimization, it is necessary to verify the compression effect and evaluate the reconstruction quality. The first step is to reconstruct the image data. The reconstruction process essentially maps data from the low-dimensional feature space back to the original high-dimensional image space, using matrix multiplication to achieve this transformation: X̂ = YV T Where X̂ is the reconstructed image data matrix, Y is the feature coefficient matrix, and V is the principal component vector matrix. Considering the special characteristics of medical ultrasound images, a reconstruction algorithm optimized for sparse data is implemented to avoid processing a large number of zero-value elements. Specifically, a block reconstruction strategy is adopted, dividing the original image space into multiple sub-blocks, and performing reconstruction calculations only on sub-blocks containing non-zero elements, significantly improving computational efficiency. For periodic sequences such as cardiac ultrasound, the system further utilizes the redundancy of the time dimension, first reconstructing keyframes (such as the endpoint frames of the cardiac cycle), and then generating intermediate frames based on these keyframes and corresponding feature coefficients through interpolation, further accelerating the reconstruction process. During the reconstruction process, an adaptive precision control mechanism is also applied, allocating computational resources according to the clinical importance of different regions: for diagnostically critical regions (such as the endocardial boundary and valve region in cardiac ultrasound), a high-precision reconstruction algorithm is used; for background regions, a simplified algorithm is used, improving computational efficiency while ensuring visual quality. After reconstruction, the system performs post-processing optimization, including edge enhancement, noise suppression, and contrast adjustment, to further improve the visual quality and diagnostic value of the reconstructed images. These optimizations are specifically designed for the characteristics of ultrasound images, such as adaptive filtering for speckle noise and local sharpening for blurred boundaries. This efficient and targeted reconstruction process enables the recovery of high-quality medical images from compressed features, laying the foundation for subsequent quality assessment and clinical applications.

[0211] Calculate the compression ratio index and evaluate the degree of storage space saving and transmission efficiency improvement by comparing the data storage size before and after compression.

[0212] After image reconstruction, the technical benefits of compression are first evaluated, namely storage space savings and improved transmission efficiency. Multi-level compression ratio metrics are calculated to comprehensively assess the compression effect. First, the original data compression ratio (OCR) is calculated: OCR = Original data size / Compressed data size, where the compressed data includes the characteristic coefficient matrix Y and the principal component matrix V. For typical ultrasound sequences, the OCR is usually between 4:1 and 10:1, depending on the data characteristics and the number of principal components retained. Second, the effective information compression ratio (EICR) is calculated: EICR = Non-zero element data size / Compressed data size. This metric more accurately reflects the compression effect on actual effective information, excluding the influence of a large number of zero values ​​in the original data. Third, the variable compression ratio (VCR) is calculated, which is a weighted compression ratio based on the ROI (Region of Interest): VCR = Σ(w i× Original size of region i) / Σ(w i × (size of region i after compression), where w i The system assigns weights to regions to reflect clinical importance. For example, in echocardiography, the cardiac chambers and myocardium are weighted higher than the background region. The system also evaluates the improvement in transmission efficiency due to compression, calculating the Theoretical Transmission Time Reduction Ratio (TTTR): TTTR = (Original Data Size / Network Bandwidth) / (Compressed Data Size / Network Bandwidth) = OCR. To verify the actual effect, the system conducts transmission tests under different network conditions, recording the actual transmission time and calculating the Actual Transmission Time Reduction Ratio (ATTR). Typically, ATTR is slightly lower than TTTR due to protocol overhead and network fluctuations during transmission. The system also evaluates the impact of compression on the storage system, calculating the Storage Cost Saving Rate (SCSR): SCSR = (1 - 1 / OCR) × 100%, reflecting the percentage reduction in storage costs resulting from compression. These multi-dimensional compression benefit metrics provide a comprehensive technical evaluation of the system, helping to determine the practical value and applicable scenarios of the compression strategy.

[0213] Assess the retention of key diagnostic information by comparing the contrast of lesion areas, edge sharpness, and the degree of preservation of texture details in images before and after compression.

[0214] Beyond technical specifications, the core value of medical image compression lies in preserving key diagnostic information. The system implements a diagnostic information retention assessment specifically tailored to the characteristics of ultrasound images. First, it assesses lesion region contrast retention (LCRP): LCRP = (compressed lesion vs. background contrast) / (original lesion vs. background contrast) × 100%, with an ideal value of no less than 95%. This is achieved using an automatic segmentation algorithm or referencing physician-annotated ROIs to calculate the contrast ratio between the lesion region and surrounding tissues. For different types of ultrasound examinations, different lesion characteristics are considered: cardiac ultrasound focuses on the contrast of areas with abnormal ventricular wall motion; abdominal ultrasound focuses on the boundary contrast of masses and cysts; and vascular ultrasound focuses on the contrast between plaques and vessel walls. Second, it assesses edge sharpness retention (ESPR): ESPR = (compressed edge sharpness) / (original edge sharpness) × 100%, achieved by comparing the gradient amplitude of key anatomical structure edges. The system focuses on assessing the sharpness retention of critical diagnostic boundaries, such as the endocardial boundary, organ contours, and vessel wall boundaries. Third, texture detail retention (TDPR) is evaluated: TDPR = (cosine similarity between compressed and original texture feature vectors) × 100%, achieved by extracting texture descriptors such as gray-level co-occurrence matrix features and local binary pattern features. Texture information is crucial for determining tissue properties, such as liver echo texture assessment and myocardial texture analysis. Dynamic information retention (DIPR) is also evaluated, focusing on the temporal dimension of the ultrasound sequence: DIPR = (similarity between compressed and original motion trajectories) × 100%, achieved by tracking the motion trajectories of key anatomical landmarks. For example, dynamic features such as ventricular wall motion and valve opening and closing motion in cardiac ultrasound. Finally, comprehensive diagnostic information retention (CDIR) is calculated: CDIR = W1 ×LCRP + W2 ×ESPR + W3 ×TDPR + W4 ×DIPR, with weighting coefficients optimized for different examination types. This multi-dimensional diagnostic information assessment ensures that the compression process preserves the visual features in ultrasound images that are crucial for clinical diagnosis.

[0215] Establish an image quality evaluation system, including objective evaluation indicators such as peak signal-to-noise ratio, structural similarity index, and visual information fidelity.

[0216] In addition to specific diagnostic information assessment, a comprehensive image quality evaluation system was established, employing multiple complementary objective evaluation indicators. First, the peak signal-to-noise ratio (PSNR) was calculated: PSNR = 10·log 10 (MAX 2The mean squared error (MSE) is calculated as follows: PSNR is the most basic distortion measure, reflecting pixel-level reconstruction accuracy. For ultrasound images, a baseline PSNR threshold of 35 dB is set; values ​​below this may affect diagnostic quality. However, PSNR cannot fully reflect structural information preservation, therefore, the Structural Similarity Index (SSIM) is further calculated. SSIM assesses image similarity by combining brightness, contrast, and structure, with a value ranging from 0 to 1; values ​​closer to 1 indicate better structural preservation. Spatially weighted SSIM calculations are performed, assigning higher weights to critical diagnostic regions to generate a weighted SSIM (wSSIM) index. For ultrasound images, wSSIM is required to be no less than 0.92. Third, Visual Information Fidelity (VIF) is calculated. This is an information theory-based evaluation metric that measures the degree of visual information preservation from a reference image to a distorted image. VIF is particularly suitable for evaluating textured medical images and better reflects the perception of the human visual system. The system also calculates the Feature Similarity Index (FSIM), which combines phase consistency and gradient magnitude information to better assess the preservation of high-level features such as edges and textures. For dynamic sequences, a temporal dimension is added for evaluation, calculating Temporal SSIM (T-SSIM) and Temporal VIF (T-VIF) to assess the quality consistency and temporal information preservation between consecutive frames. The system also implements a Perception-Based Image Quality Evaluator (PIQE) and a Referenceless Image Spatial Quality Evaluator (NIQE). These referenceless evaluation methods do not require the original image and are suitable for scenarios lacking complete references. Finally, the system calculates the Comprehensive Objective Quality Score (COQS): COQS = w1×norm(PSNR) + w2×SSIM + w3×VIF + W4 ×FSIM + W5 ×T-SSIM, where norm() is the normalization function that maps PSNR to the 0-1 range. This multi-indicator fusion evaluation system provides a comprehensive quality assessment, overcoming the limitations of a single indicator.

[0217] Based on the compression ratio index, the retention rate of key diagnostic information, and the objective evaluation index, an image quality index is constructed.

[0218] To comprehensively evaluate the overall quality of compressed ultrasound images, a multi-level fusion image quality index system was constructed. First, the Analytic Hierarchy Process (AHP) was used to determine the weights of each evaluation dimension: technical benefit dimension (including compression ratio), diagnostic value dimension (including retention of key diagnostic information), and visual quality dimension (including objective evaluation indicators). The weight allocation was based on clinical expert questionnaires and extensive case analysis; typically, the diagnostic value dimension had the highest weight (approximately 50-60%), followed by the visual quality dimension (approximately 25-35%), and the technical benefit dimension had the lowest weight (approximately 10-15%). Within each dimension, a weighted geometric mean method was used to integrate sub-indicators. This method is more sensitive to extremely low values, ensuring that a significant decrease in any key indicator will significantly affect the final score. For example, the diagnostic value dimension score DV = (LCRP) / (LCRP / (LCRP)). W1 × ESPR W2 ×TDPR W3 × DIPR W4 ) 1 / Σwi The weights are dynamically adjusted based on the examination type. The system also implements an adaptive evaluation mechanism, adjusting the evaluation strategy according to different examination types and clinical application scenarios: for cardiac ultrasound, DIPR (Detailed Information Retention) has a higher weight; for abdominal ultrasound, TDPR (Texture Detail Retention) has a higher weight; and for vascular ultrasound, ESPR (Edge Sharpness Retention) has a higher weight. Finally, the Comprehensive Quality Index (CIQI) is calculated.

[0219]

[0220] TE, DV, and VQ are the scores for the three dimensions of technical benefits, diagnostic value, and visual quality, respectively, with w representing the corresponding weight. CIQI values ​​range from 0 to 100, with higher scores indicating better compression quality. CIQI is divided into five quality levels: Excellent (90-100), Good (80-90), Acceptable (70-80), Borderline (60-70), and Unacceptable (<60). To improve the interpretability of the evaluation, a detailed quality analysis report is generated, including scores for each dimension, key sub-indicator values, quality level, and specific descriptions of advantages and disadvantages, helping users fully understand the compression quality status.

[0221] A quality threshold is set according to clinical diagnostic needs. When the image quality index is equal to or greater than the set quality threshold, the compression quality is confirmed to meet the diagnostic requirements, and the compressed feature data is generated.

[0222] After a comprehensive quality assessment, the system proceeds to the final decision-making and data generation phase. Different quality thresholds are set based on various clinical application scenarios to ensure that the compressed data meets specific diagnostic needs. For primary diagnostic purposes (such as initial diagnosis and analysis of important cases), strict quality thresholds are set, typically requiring a CIQI ≥ 85, to ensure that compression does not affect diagnostic accuracy. For auxiliary diagnostic purposes (such as case review and teaching demonstrations), the system sets moderate quality thresholds, typically requiring a CIQI ≥ 75. For reference purposes (such as quick previews and medical history recordings), the system sets relatively lenient quality thresholds, typically requiring a CIQI ≥ 65. In addition to the overall index thresholds, the system also sets hard constraints on key sub-indicators: regardless of the CIQI score, if certain key indicators (such as LCRP and ESPR) are below the safety threshold, they will still be marked as "not recommended for diagnosis." When the assessment results meet the quality threshold requirements, the final compressed feature data package is generated, containing three parts: the compressed data core, including the feature coefficient matrix Y, the principal component matrix V, and necessary reconstruction parameters; quality assessment metadata, recording the values ​​of various quality indicators, assessment results, and applicable scenario recommendations; and technical metadata, recording compression parameters, algorithm version, and processing history. Digital signatures are applied to compressed data packets to ensure data integrity and verifiable origin. To support flexible applications, the system offers multi-level compression options, allowing users to select versions with different compression ratios and quality levels according to their needs. The system also implements a progressive transmission mechanism, allowing for the transmission of low-quality preview versions first, followed by the transmission of full-quality versions as needed, improving network utilization efficiency. Finally, the generated compressed feature data is stored in a local database, and index information is updated, recording the mapping relationship between the original and compressed data, compression ratios, and quality ratings, providing a basis for subsequent data management and use. Through this rigorous quality control and flexible data generation mechanism, the system ensures that compressed data meets both storage and transmission efficiency requirements while maintaining the image quality needed for clinical diagnosis.

[0223] Example 12

[0224] Based on Example 8, the comparison of the connected component analysis results with a preset integrity threshold, and the marking of any connected component with a connectivity degree lower than the preset threshold as an abnormal sequence requiring reconstruction, and the generation of a sequence quality assessment report, includes:

[0225] Establish multi-level quality assessment standards, including pixel-level grayscale continuity assessment, feature-level anatomical structure consistency assessment, and semantic-level physiological parameter rationality assessment.

[0226] When evaluating the quality of ultrasound image sequences, a hierarchical quality assessment framework is employed, comprehensively analyzing sequence integrity from low-level pixel information to high-level semantic understanding. At the pixel-level assessment level, the system focuses on the continuity of basic image characteristics, which is the most fundamental guarantee of sequence integrity. The system implements an adaptive gray-level continuity assessment algorithm, quantifying pixel-level continuity by calculating the gray-level histogram similarity between adjacent frames, the rate of gray-level change in local regions, and the consistency of overall brightness distribution. Different assessment parameters are used for different types of ultrasound examinations: for cardiac ultrasound, considering the periodic changes in myocardial motion, the system uses an adaptive threshold based on the cardiac cycle; for abdominal ultrasound, more attention is paid to the gray-level transition characteristics of tissue interfaces; for vascular ultrasound, the focus is on assessing the continuity of echoes within the vascular lumen. At the feature-level assessment level, the spatial continuity and morphological consistency of anatomical structures are considered, which are crucial for clinical diagnosis. A deep learning-based anatomical structure recognition algorithm is implemented to automatically locate key anatomical landmarks (such as endocardial boundary points and valve attachment points in cardiac ultrasound), track the movement trajectories of these landmarks in the sequence, and assess their spatial continuity. The system also analyzes the deformation characteristics of organ contours, detecting abnormal morphological changes or jumps. At the semantic level, verifying the physiological rationality of the sequence from a medical diagnostic perspective represents the highest level of quality assurance. A knowledge graph-based physiological parameter verification mechanism was implemented to extract key physiological parameters (such as ejection fraction and wall motion amplitude in cardiac ultrasound, and blood flow velocity in vascular ultrasound), analyze the temporal trends of these parameters, and verify whether they conform to physiological laws. Through this multi-level quality assessment standard, the integrity and reliability of ultrasound imaging sequences can be comprehensively and deeply analyzed, providing a solid guarantee for clinical diagnosis.

[0227] Quantitative analysis of pixel-level quality is performed, and statistics of grayscale difference between adjacent frames, texture correlation coefficient, and motion vector consistency index are calculated.

[0228] In pixel-level quality analysis, a series of refined statistical analysis methods were implemented to comprehensively evaluate the basic image characteristics of the sequence. First, gray-level difference statistics between adjacent frames were calculated, including mean squared error (MSE), mean absolute difference (MAD), and structural similarity index (SSIM). To improve analysis accuracy, a region-adaptive strategy was adopted, dividing the image into multiple regions of interest (ROIs) and calculating weighted difference statistics for different regions. For example, for cardiac ultrasound, the heart chambers and myocardium regions have higher weights than the background region; for abdominal ultrasound, the target organ region has a higher weight than the surrounding tissue. Histogram-based difference analysis was also implemented, calculating the Bhattacharyya distance, KL divergence, and Earth Mover's distance between the gray-level histograms of adjacent frames. These indicators can more comprehensively reflect changes in gray-level distribution. Second, texture correlation coefficients were calculated to assess the continuity of tissue texture features. The system extracts multi-scale, multi-directional texture features, including gray-level co-occurrence matrix (GLCM) features, local binary pattern (LBP) features, and Gabor filter features, and calculates the correlation coefficients of these features between adjacent frames. For the speckle noise unique to ultrasound images, specialized speckle correlation analysis was implemented to assess the temporal consistency of speckle patterns, which is crucial for tissue tracking and motion analysis. Third, motion vector consistency indices were calculated to assess the smoothness and coherence of object motion within the sequence. Pixel-level motion vector fields were estimated using block matching and optical flow methods, analyzing the directional consistency, amplitude continuity, and temporal smoothness of motion vectors. Abnormal motion jumps, such as sudden large displacements or directional reversals, were detected, which typically indicate sequence problems. To accommodate different types of physiological motion, multi-modal motion analysis was implemented: for periodic motion (such as heartbeats), the periodic consistency of motion was systematically assessed; for quasi-static scans (such as abdominal organ scans), the progressiveness and directional consistency of motion were assessed. Through these refined pixel-level analyses, minute abnormalities in the sequence can be accurately identified, providing a solid foundation for high-level quality assessment.

[0229] The spatial continuity of anatomical landmarks is verified at the feature level, and the geometric consistency of organ contours, blood vessel orientation, and tissue interfaces is detected.

[0230] In feature-level quality assessment, the spatial continuity and morphological consistency of key anatomical structures in ultrasound images are crucial for clinical diagnosis. First, an automated anatomical landmark detection and tracking algorithm was implemented. Based on deep learning models (such as U-Net and Faster R-CNN), key anatomical landmarks were identified, such as endocardial and epicardial boundaries, valve opening and closing points, and papillary muscle positions in cardiac ultrasound; organ boundaries and vascular bifurcation points in abdominal ultrasound; and vessel wall boundaries and plaque boundaries in vascular ultrasound. The motion trajectories of these landmarks in the sequence were tracked, and smoothness indices (such as rate of curvature change and acceleration consistency) and continuity indices (such as position jump detection and missing point ratio) were calculated. For periodically moving landmarks, the system verified the closure and repeatability of their periodic trajectories; for non-periodic moving landmarks, the system verified the physical rationality of their motion trajectories. Second, the geometric consistency of organ contours was assessed. Organ boundaries were extracted using an active contour model or level set method, and the temporal evolution characteristics of the boundary shapes were analyzed. The system calculates the rate of change of shape descriptors (such as Fourier descriptors and moment invariants) over time to detect anomalous shape changes or jumps. It also implements shape constraint verification based on anatomical knowledge; for example, the heart chambers should maintain specific morphological proportions during systole, and the liver contour should maintain smooth, continuous curve characteristics. Third, the system analyzes the geometric consistency of vascular orientation and tissue interfaces, extracting linear structures through Hessian matrix analysis or directional filters, and tracking changes in the vascular centerline and tissue interfaces throughout the sequence. The system calculates the directional continuity, curvature changes, and topological stability of these structures, detecting anomalous breaks or deformations. For important tissue interfaces (such as the myocardium-endocardium interface and liver-kidney interface), specialized interface stability analysis is implemented to assess the temporal consistency of interface position and clarity. Through these feature-level analyses, sequence quality can be evaluated from an anatomical perspective, identifying structural discontinuities or morphological abnormalities that may affect diagnosis.

[0231] The clinical rationality of the sequence was verified by analyzing the trends of physiological parameter changes, including testing the reasonable range of physiological indicator changes and tissue movement amplitude.

[0232] In semantic-level quality assessment, verifying the physiological rationality of ultrasound sequences from a medical diagnostic perspective represents the highest level of quality assurance. First, an automated physiological parameter extraction algorithm was implemented to calculate key diagnostic indicators from ultrasound sequences: for cardiac ultrasound, parameters such as left ventricular volume, ejection fraction (EF), myocardial wall thickness, and valve opening and closing amplitude were calculated; for vascular ultrasound, parameters such as vessel diameter, blood flow velocity, and resistance index were calculated; and for abdominal ultrasound, parameters such as organ size, parenchymal echo intensity, and the proportion of cystic structures were calculated. The temporal trends of these parameters were analyzed to verify their conformity to physiological laws. For example, cardiac ejection fraction should remain relatively stable within a cardiac cycle, and blood flow velocity should exhibit a pulsatile pattern synchronized with the cardiac cycle. A parameter rationality verification mechanism based on a medical knowledge base was established, setting normal ranges and rate-of-change limits for various physiological parameters and marking abnormal values ​​exceeding the reasonable range. Second, the rationality of tissue motion amplitude was assessed, verifying the motion characteristics of different tissue structures based on anatomical knowledge. For example, the contraction amplitude of the myocardium should conform to the range of normal cardiac function, and organ displacement caused by respiratory movements should exhibit regular periodic changes. This system achieves motion pattern analysis based on statistical models, learns the characteristic distribution of normal physiological motion, and detects abnormal motions that deviate from normal patterns. Third, it verifies the synergistic relationships between multiple parameters and checks the logical consistency between different physiological indicators. For example, increased cardiac wall motion during systole should be synchronized with a decrease in cardiac chamber volume, and an increase in vessel diameter should show a specific relationship with changes in blood flow velocity. A physiological parameter association network is established, and conditional dependencies between parameters are analyzed based on Bayesian networks or graphical models to identify abnormal patterns that violate physiological laws. Finally, the system considers the influence of individual patient factors on physiological parameters, adjusting parameter evaluation criteria according to the patient's age, gender, body type, and clinical condition to achieve personalized rationality verification. Through this semantic-level analysis based on medical knowledge, sequence quality can be assessed from a clinical diagnostic perspective, ensuring the medical reliability of ultrasound examination results.

[0233] Generate a sequence quality assessment report that includes quality score, abnormal location markers, and repair suggestions, providing clinicians with a basis for evaluating the reliability of imaging sequences and for diagnostic reference.

[0234] After completing the multi-level quality assessment, all analysis results are integrated to generate a comprehensive and intuitive sequence quality assessment report, providing decision support for clinicians. The report adopts a hierarchical structure, including three main parts: summary, detailed assessment, and recommendations. The summary provides an overview of the overall sequence quality: a comprehensive quality score, which integrates pixel-level, feature-level, and semantic-level assessment results based on a weighted fusion algorithm, with a score range of 0-100, divided into five levels: "Excellent (90-100)," "Good (80-90)," "Average (70-80)," "Poor (60-70)," and "Unusable (<60)"; a quality radar chart, which intuitively displays the assessment results of each dimension, including six dimensions: grayscale continuity, texture consistency, motion smoothness, anatomical continuity, morphological consistency, and physiological rationality; and an anomaly overview, which concisely lists the main problems detected, sorted by severity. In the detailed assessment section, the system provides multi-level quality analysis results: pixel-level assessment details, including numerical values ​​of various statistical indicators, time-varying curves, and heatmaps of abnormal regions; feature-level assessment details, including trajectory maps of key anatomical landmarks, morphological change analysis of organ contours, and geometric consistency scores; and semantic-level assessment details, including time-series plots of physiological parameters, normal range markings, and outlier analysis. The system provides a detailed description of each detected abnormality, including the abnormality type, location (frame index and spatial region), severity, and possible cause. In the recommendations section, the system provides targeted repair and usage suggestions: for minor issues, automatic repair options are provided, such as interpolation smoothing, noise suppression, or small-area repair; for moderate issues, repair suggestions and precautions are provided, such as "moderate breaks exist between frames 45-48; periodic interpolation is recommended for repair, and the continuity of myocardial motion should be verified after repair"; for severe issues, clear usage limitations are provided, such as "the sequence has severe quality problems and is not recommended for quantitative analysis; it can only be used for qualitative reference." The system also provides suitability assessments for specific clinical applications, such as "The current sequence is suitable for measuring cardiac chamber volume, but not for assessing regional wall motion abnormalities." The report features an interactive design, allowing medical staff to click on abnormal areas to view detailed analysis, adjust display parameters, and even initiate automatic repair processes. This comprehensive, detailed, and practical quality assessment report provides clinicians with an objective evaluation and clear guidance on sequence quality, ensuring that diagnoses are based on reliable data and improving the diagnostic value and reliability of ultrasound examinations.

[0235] Example 12

[0236] Based on Example 4, the application of a pre-trained deep learning model to the multidimensional feature space for lesion region detection, the identification of abnormal tissue regions through a convolutional neural network and the calculation of confidence scores to obtain lesion region detection results, specifically includes:

[0237] During the training data preparation phase, a training dataset of ultrasound images was collected and organized, including 20,000 annotated ultrasound images. This dataset contains 12,000 normal tissue samples and 8,000 samples with different types of lesions (3,000 cysts, 2,500 solid masses, 1,500 calcifications, and 1,000 mixed lesions). The training data originated from the ultrasound departments of three top-tier hospitals and underwent independent annotation and cross-validation by three senior ultrasound physicians to ensure annotation quality. All data underwent privacy protection processing, removing patient-identified information.

[0238] In the data preprocessing stage, the original ultrasound images were standardized: first, the images were adjusted to a uniform resolution (512×512 pixels); then, adaptive histogram equalization was applied to enhance contrast; next, anisotropic diffusion filtering was used to reduce noise while preserving edge information; finally, grayscale normalization was performed, mapping pixel values ​​to the 0-1 range. To enhance the model's generalization ability, data augmentation strategies were also implemented, including random rotation (±15°), translation (±10%), scaling (0.9-1.1 times), elastic deformation, and random noise addition, expanding the training set to three times its original size.

[0239] The model architecture design adopts an improved U-Net structure, specifically including: an encoding path containing 5 encoding blocks, each consisting of two 3×3 convolutional layers (padding 1, stride 1), a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer (stride 2), with the number of convolutional kernels starting from 64 and doubling with each layer up to 1024; a decoding path containing 4 decoding blocks, each consisting of a 2×2 deconvolutional layer (stride 2), a concatenation layer (connected to the corresponding encoded feature map), two 3×3 convolutional layers (padding 1, stride 1), a batch normalization layer, and a ReLU activation function, with the number of convolutional kernels halved with each layer starting from 512; and an output layer that uses 1×1 convolutions to convert the feature map into a prediction mask, followed by a Sigmoid activation function to output a probability map in the range of 0-1.

[0240] To improve the model's ability to identify small lesions, a spatial attention mechanism is introduced, which enhances the feature representation of important regions by learning spatial weight maps. Specifically, an attention module is added at the encoder-decoder bridge, including: feature map compression (reducing the number of channels to 1 / 8 of the original through 1×1 convolution), spatial feature extraction (through two 3×3 convolutional layers), attention weight generation (through 1×1 convolution and the Sigmoid activation function), and feature map weighting (multiplying the original feature map with the attention weight map).

[0241] The model training used the following hyperparameter settings: batch size of 16; initial learning rate of 0.001, dynamically adjusted using cosine annealing; Adam optimizer with parameters β1=0.9, β2=0.999, and ε=1e-8; a weighted combination loss function, including binary cross-entropy loss and Dice loss, with a weight ratio of 0.4:0.6; 200 training epochs, stopping early when the validation set performance no longer improved; a weight decay coefficient of 1e-5 to mitigate overfitting; and a learning rate warm-up strategy introduced after the 100th epoch, increasing the learning rate to 1.2 times its original value every 10 epochs before further decaying to escape local optima.

[0242] Model validation and evaluation employed 5-fold cross-validation. Evaluation metrics included: sensitivity (mean 92.7%), specificity (mean 94.3%), Dice coefficient (mean 89.5%), Jaccard index (mean 83.2%), and AUROC (mean 0.951). Model performance varied slightly across different lesion types: cystic lesions showed the best detection performance (Dice coefficient 92.3%), followed by solid masses (Dice coefficient 88.7%), while calcifications and mixed lesions showed relatively lower performance (Dice coefficients 86.5% and 85.8%, respectively). Blind comparison with three senior ultrasound physicians showed that the model's detection sensitivity was close to expert levels (model 92.7% vs. expert average 94.1%), but there was still a gap in specificity (model 94.3% vs. expert average 97.8%).

[0243] Model deployment optimization mainly includes two aspects: 1) Model compression: the original model (about 47MB) is compressed into a lightweight model (about 12MB) through knowledge distillation, with the accuracy loss controlled within 3%; 2) Computation acceleration: through ONNX format conversion and quantization technology (INT8 quantization), the inference time of the model on ordinary CPU devices is reduced from the original 1.2 seconds to 0.3 seconds, meeting the real-time processing requirements in offline environments.

[0244] Example 13

[0245] Based on Example 5, the steps for optimizing the synchronization strategy using the maximum weighted independent set algorithm based on the sparse graph structure are further optimized, including:

[0246] A Congest approximation algorithm based on graph theory for the maximum weighted independent set in sparse graphs is implemented. This algorithm is particularly suitable for complex constraints in medical image data synchronization scenarios.

[0247] The algorithm input includes: G=(V,E) represents the sparse graph structure, where V is the set of data item nodes and E is the set of dependency / conflict relationship edges; w: V→R+ represents the node weight function, reflecting the data synchronization priority; d: V→N represents the node degree function, reflecting the number of constraint relationships of data items.

[0248] The algorithm execution process is as follows:

[0249] Initialization: Mark all nodes as "unprocessed", set the independent set I=∅, set the current round t=1, and set the maximum round tmax=log|V|;

[0250] For each round t∈[1,tmax], perform the following steps:

[0251] (a) Message passing phase: Each unprocessed node v sends (ID(v), w(v)) to all its neighbors;

[0252] (b) Local computation phase: Each unprocessed node v receives messages from its neighbors and computes:

[0253] - Unprocessed neighbor set N'(v) = {u∈N(v): u has not been processed};

[0254] - The decision value DV(v) = w(v) / (|N'(v)|+1)α, where α is an adjustment parameter with a default value of 0.5;

[0255] (c) Probabilistic selection phase: Each unprocessed node v decides whether to join the candidate independent set I' with probability p(v)=min{1, DV(v) / max{DV(u): u is a two-hop neighbor of v}};

[0256] (d) Conflict resolution phase: For a node in I', if there is an adjacent node in I', the node with the lower weight exits I';

[0257] (e) Results update phase:

[0258] - I=I∪I';

[0259] - Mark all nodes in I' and their neighbors as "processed".

[0260] Post-processing stage: Apply topological sorting to verify the dependencies of nodes in the independent set I. If there are violations of dependency constraints, perform minimum weight loss adjustment.

[0261] In the specific implementation, the following optimization techniques are introduced:

[0262] Adaptive adjustment: The α value is dynamically adjusted according to the density of the graph. A larger α value (about 0.7) is used for sparse graphs, and a smaller α value (about 0.3) is used for dense graphs.

[0263] Block processing: Divide the large-scale graph into overlapping subgraphs, perform parallel computations, and then merge the results;

[0264] Preprocessing optimization: For explicit linear dependency chains, determine the optimal order in advance;

[0265] Second-level cache: Maintains a cache of the decision values ​​of a node's two-hop neighbors to reduce redundant calculations;

[0266] Adaptive round control: Terminate early when the number of incremental nodes is lower than the threshold (|V|*0.01) for two consecutive rounds.

[0267] Algorithm performance analysis shows that the algorithm can keep the computation time within 200ms on a typical medical data synchronization graph with 1000 nodes and about 3000 edges, and the total weight of the independent sets obtained reaches about 92% of the theoretical optimal solution, which is significantly better than the greedy algorithm (about 85%) and the random algorithm (about 70%).

[0268] In practical applications, the algorithm demonstrates its workflow through a specific example: Assume that ultrasound examination data for patient Zhang San was generated during offline processing, including basic patient information (node ​​A, weight 7.5), examination records (node ​​B, weight 8.2), cardiac ultrasound DICOM sequence (node ​​C, weight 9.1), and ultrasound report (node ​​D, weight 9.8). A dependency relationship exists: A→B→C→D. Traditional sequential synchronization requires transmitting A→B→C→D in sequence, while this algorithm can identify only the dependency chain A→B→C→D and can immediately begin transmitting A, while reserving resources for subsequent nodes, significantly shortening the total synchronization time.

[0269] Example 14

[0270] like Figure 2 As shown, this embodiment provides an offline ultrasound image acquisition and dynamic synchronous processing system, including:

[0271] The network status monitoring module 201 is used to acquire network connection status data between the client and the server, continuously monitor the network connection status data and perform connectivity detection, determine the current network working mode, and obtain the network working mode result.

[0272] The online data synchronization module 202 is used to retrieve pending inspection task data and report template data from the server when the network working mode result is online, store the pending inspection task data and the report template data in a local cache database, and establish an online data synchronization mechanism.

[0273] The image acquisition and processing module 203 is used to acquire multimodal image data acquired by ultrasound equipment based on the online data synchronization mechanism, perform DICOM protocol parsing and format conversion processing on the multimodal image data to obtain converted image data, and write the converted image data to both a local temporary library and a server PACS system to generate local image backup data.

[0274] The offline mode switching module 204 is used to detect network interruption events and trigger offline mode switching based on changes in the network connection status data, call the local SQLite database to take over the data storage service, activate the local image backup data, and start the local image acquisition and processing process.

[0275] The offline diagnostic processing module 205 is used to perform local AI algorithm analysis and lesion identification processing on ultrasound image data acquired in offline state, and generate an offline diagnostic report by combining the report template data in the local cache database, thus completing the offline business closed loop.

[0276] The incremental synchronization module 206 is used to continuously monitor the network connection status data, detect network recovery events and trigger the data synchronization process, identify the data to be synchronized generated during offline periods through a difference comparison algorithm, and perform incremental data upload to the server to complete the offline acquisition and dynamic synchronization processing of multimodal fusion ultrasound images.

[0277] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An offline acquisition and dynamic synchronization processing method of ultrasonic images, characterized in that, The method comprises the following steps: obtaining network connection state data between a client and a server, continuously monitoring and detecting the network connection state data, judging a current network working mode, and obtaining a network working mode result; when the network working mode result is an online state, pulling to-be-done examination task data and report template data from the server, storing the to-be-done examination task data and the report template data to a local cache database, and establishing an online data synchronization mechanism; based on the online data synchronization mechanism, obtaining multi-modal image data collected by an ultrasonic device, performing DICOM protocol analysis and format conversion processing on the multi-modal image data, obtaining converted image data, and double-writing and storing the converted image data to a local temporary library and a server PACS system, and generating local image backup data; Based on the change of the network connection state data, a network interruption event is detected and an offline mode switching is triggered, a local SQLite database is called to take over the data storage service, the local image backup data is activated and a localized image acquisition and processing flow is started, including: constructing each frame in a dynamic ultrasound image sequence acquired in an offline state as a node in a graph structure, calculating the spatio-temporal correlation between frames as an edge weight, and establishing a time correlation graph; performing connected component analysis on the time correlation graph, identifying missing frames, damaged frames and time sequence abnormal areas in the image sequence, counting the number of connected components and topological properties, and obtaining a connected component analysis result; based on the connected component analysis result, combining the pre-processing feature library in the local image backup data, applying an optimal graph reconstruction algorithm to repair the integrity of the image sequence, and completing missing data segments through feature interpolation and periodic reconstruction methods to obtain a reconstructed image sequence; performing multi-level consistency verification on the reconstructed image sequence, including pixel-level gray continuity verification, feature-level anatomical structure continuity verification and semantic-level physiological parameter rationality verification, to generate complete localized image data; storing the complete localized image data in layers according to patient ID, examination item and image sequence to obtain the execution result of the localized image acquisition and processing flow; wherein the connected component analysis on the time correlation graph, the identification of missing frames, damaged frames and time sequence abnormal areas in the image sequence, the counting of the number of connected components and topological properties, and the obtaining of the connected component analysis result, include: based on a depth-first search algorithm, all nodes of the time correlation graph are traversed, mutually connected node sets are identified, and connected components and the number of connected components are obtained; topological property parameters of each connected component are calculated, including connectivity, diameter, clustering coefficient and centrality index, the integrity degree of the image sequence is evaluated, and an integrity result is obtained; by analyzing the time span and spatial distribution characteristics of the connected components, the sequence break position caused by device failure, transmission interruption or buffer overflow is identified; based on the physiological periodicity rule, the rationality of the connected components is verified, and for the connected mode corresponding to the physiological cycle of the periodic image, a physiological verification result is obtained; based on the number of connected components, the integrity result, the sequence break position and the physiological verification result, a connected component analysis result is constructed; the connected component analysis result is compared with a preset integrity threshold, and when the connectivity of any connected component is lower than the preset threshold, it is marked as an abnormal sequence that needs to be reconstructed, and a sequence quality evaluation report is generated; ultrasound image data acquired in an offline state is subjected to local AI algorithm analysis and lesion identification processing, and an offline diagnosis report is generated based on the report template data in the local cache database, completing an offline business closed loop; The network connection state data is continuously monitored, a network recovery event is detected, and a data synchronization process is triggered, the generated to-be-synchronized data during the offline period is identified through a difference comparison algorithm, incremental data upload to the server is performed, and offline acquisition and dynamic synchronization processing of the multi-modal fusion ultrasound image are completed, including: constructing the to-be-synchronized data into a sparse graph structure, wherein each data unit is taken as a node, and the dependency relationship and conflict relationship between the data are taken as edge connections, and the node weight value is calculated; applying a maximum weighted independent set algorithm based on the sparse graph structure to optimize the synchronization strategy, and solving a data subset that can be synchronized in parallel and has the maximum weight through a distributed approximation algorithm; processing the data subset in batches according to the dependency relationship, preferentially synchronizing emergency data and basic dependency data with high weight values, to avoid data conflicts and transmission failures; in a multi-client environment, a synchronization plan is coordinated through a lightweight coordination protocol to prevent server load impact and optimize network resource utilization efficiency; performing state write-back and local cleaning operations on the synchronized data, marking the synchronized data and cleaning up temporary storage, to obtain the execution result of the incremental data upload to the server.

2. The method of claim 1, wherein, Based on the online data synchronization mechanism, multi-modal image data acquired by an ultrasound device is obtained, DICOM protocol analysis and format conversion processing are performed on the multi-modal image data, converted image data is obtained, and the converted image data is double-written and stored to a local temporary library and a server PACS system, to generate local image backup data, including: A three-dimensional sparse tensor structure is constructed for the multi-modal image data, wherein the first dimension represents a spatial position coordinate, the second dimension represents a time sequence, and the third dimension represents a modal feature, a region of interest is extracted, and a sparse tensor representation is generated; Based on the sparse tensor representation, a local search principal component analysis algorithm is applied for feature dimension reduction processing, an optimal principal component direction is found through gradient projection and soft thresholding operation, and compressed feature data is generated; The compressed feature data is subjected to quality evaluation, variance explanation rate and diagnostic key information retention rate are calculated, the compressed effectiveness is confirmed when the variance explanation rate reaches a preset effectiveness threshold, and the compressed image data is stored to the local temporary library; Based on the compressed feature data, image statistical features, texture descriptors, and morphological parameters are calculated, a preprocessing feature library required for offline AI analysis is generated, and the local image backup data is obtained.

3. The method of claim 1, wherein, The ultrasound image data acquired in the offline state is subjected to local AI algorithm analysis and lesion identification processing, an offline diagnosis report is generated in combination with the report template data of the local cache database, and an offline business closed loop is completed, including: Image feature vectors are extracted from the ultrasound image data acquired in the offline state, including gray histogram features, texture features, and shape features, and a multi-dimensional feature space is constructed; A pre-trained deep learning model is applied to the multi-dimensional feature space for lesion region detection, an abnormal tissue region is identified through a convolutional neural network, and a confidence score is calculated, to obtain a lesion region detection result; Perform quantitative measurement analysis based on the lesion area detection result, calculate geometric parameters such as lesion area, perimeter, long diameter and short diameter, and generate structured measurement data; Match and fuse the structured measurement data with the report template data in the local cache database, automatically fill in the diagnostic parameters and suggestion description, and generate an offline diagnostic report containing text, image annotation and AI analysis parameters; Perform integrity check and format standardization processing on the offline diagnostic report to ensure compliance with medical report specifications, and obtain the completion status of the offline business closed loop.

4. The method of claim 2, wherein, Based on the sparse tensor representation, apply local search principal component analysis algorithm for feature dimension reduction processing, find the optimal principal component direction through gradient projection and soft threshold operation, and generate compressed feature data, including: Preprocess the sparse tensor representation, extract the position and value of the non-zero elements in the medical image, and construct a sparse matrix representation form; Based on the sparse matrix representation form, apply an iterative optimization algorithm to calculate the eigenvalues and eigenvectors of the covariance matrix, find the principal component direction under the constraint of sparsity, and obtain the principal component direction; Perform soft thresholding on the principal component direction, set a threshold parameter, and set the coefficients less than the threshold parameter to zero to maintain sparsity, and obtain the principal component coefficients; Update the principal component coefficients through gradient projection method to ensure that the projection result meets the sparsity constraint condition, and iterate and optimize until convergence to obtain the optimized principal component coefficients; Based on the optimized principal component coefficients, reconstruct the compressed feature data, calculate the compression ratio and feature retention degree evaluation index, and obtain the compressed feature data.

5. The method of claim 3, wherein, Based on the lesion area detection result, perform quantitative measurement analysis, calculate geometric parameters such as lesion area, perimeter, long diameter and short diameter, and generate structured measurement data, including: Perform boundary extraction processing on the lesion area detection result, apply edge detection algorithm to identify the lesion contour and perform morphological filtering and denoising; Calculate the area parameter based on the lesion contour, obtain the number of pixels in the lesion area through pixel statistical method, and convert to actual area value combined with pixel spacing; Measure the perimeter of the lesion contour, calculate the perimeter of the lesion contour using contour tracking algorithm, and apply curve smoothing algorithm to eliminate sawtooth effect; Calculate the long diameter and short diameter of the lesion contour by the minimum circumscribed rectangle method, determine the main direction and geometric parameters of the lesion contour; Calculate the shape descriptor based on the geometric parameters, including circularity, thin length and irregularity, and construct the structured measurement data containing position, size, shape and texture features.

6. The method of claim 1, wherein, Apply the maximum weighted independent set algorithm to the sparse graph structure to optimize the synchronization strategy, solve the data subset with maximum weight that can be synchronized in parallel through distributed approximation algorithm, including: Calculate the priority weight of the data to be synchronized using a multi-factor weighting method, including data priority factor, time urgency factor and data size influence factor; Based on the data dependency relationship, construct the edge connection of the graph, including the association dependency between the report and the image, the mapping dependency between the patient ID and the examination data, and the conflict constraint of the device resources; A distributed algorithm is applied for multi-round iterative computation, and in each round, each node updates its state based on local information and coordinates global optimization through a message passing mechanism; In each iteration, a candidate set of computing nodes is calculated, and the data item with the maximum weight and no conflict with the selected nodes is selected to join the independent set until it cannot be expanded; The feasibility of the maximum weighted independent set obtained by calculation is verified to ensure that the selected data items meet the parallel synchronization conditions and do not violate the dependency constraints, and the optimal synchronization strategy scheme is generated.

7. The method of claim 4, wherein, Based on the optimized principal component coefficients, the compressed feature data is reconstructed, the compression ratio and feature retention evaluation index are calculated, and the compressed feature data is obtained, including: Based on the principal component coefficients, the image data is reconstructed, and the compressed features are mapped back to the original image space through matrix multiplication operation; Calculate the compression ratio index, compare the data storage size before and after compression, and evaluate the storage space saving degree and transmission efficiency improvement effect; Evaluate the retention degree of diagnostic key information, compare the contrast, edge definition and texture detail retention degree of the lesion area of the image before and after compression; An image quality evaluation system is established, including objective evaluation indexes such as peak signal-to-noise ratio, structural similarity index and visual information fidelity; Based on the compression ratio index, the diagnostic key information retention degree and the objective evaluation index, an image quality index is constructed; According to the clinical diagnosis requirements, set the quality threshold, when the image quality index is equal to or greater than the set quality threshold, confirm that the compression quality meets the diagnostic requirements, and generate the compressed feature data.

8. The method of claim 1, wherein, The connectivity component analysis result is compared with a preset integrity threshold, and when the connectivity of any of the connected components is lower than the preset threshold, the abnormal sequence needing reconstruction is marked, and a sequence quality evaluation report is generated, including: A multi-level quality evaluation standard is established, including pixel-level gray level continuity evaluation, feature-level anatomical structure consistency evaluation and semantic-level physiological parameter rationality evaluation; Quantitative analysis of pixel-level quality, calculation of gray difference statistics, texture correlation coefficient and motion vector consistency index between adjacent frames; Verify the spatial position continuity of anatomical landmark points at the feature level, detect the geometric consistency of organ contour, blood vessel direction and tissue interface; Through physiological parameter change trend analysis, verify the clinical rationality of the sequence, including physiological index change and reasonable range test of tissue motion amplitude; Generate the sequence quality evaluation report containing quality score, abnormal position mark and repair suggestion, provide image sequence reliability evaluation and diagnostic reference for clinicians.

9. An ultrasonic image off-line acquisition and dynamic synchronization processing system, characterized by comprising: an ultrasonic image off-line acquisition device; a dynamic synchronization processing device; and a synchronization processing device. Including: A network state monitoring module is configured to obtain network connection state data between a client and a server, continuously monitor and detect the network connection state data, determine a current network working mode, and obtain a network working mode result; An online data synchronization module is configured to, when the network working mode result is an online state, pull to-be-done inspection task data and report template data from the server, store the to-be-done inspection task data and the report template data in a local cache database, and establish an online data synchronization mechanism; An image acquisition and processing module is configured to acquire multi-modal image data collected by an ultrasound device based on the online data synchronization mechanism, perform DICOM protocol analysis and format conversion processing on the multi-modal image data, obtain converted image data, and write the converted image data to a local temporary library and a server PACS system to generate local image backup data; An offline mode switching module is configured to detect a network interruption event and trigger offline mode switching based on changes in the network connection state data, call a local SQLite database to take over data storage services, activate the local image backup data, and start a localized image acquisition and processing procedure, including: constructing each frame in a dynamic ultrasound image sequence collected in an offline state as a node in a graph structure, calculating inter-frame spatio-temporal correlation as an edge weight, and establishing a time-correlation graph; performing connected component analysis on the time-correlation graph to identify missing frames, damaged frames, and time-correlation abnormal regions in the image sequence, count the number of connected components and topological properties, and obtain a connected component analysis result; based on the connected component analysis result, combining a pre-processing feature library in the local image backup data, applying an optimal graph reconstruction algorithm to repair the integrity of the image sequence, and using feature interpolation and periodic reconstruction methods to complete missing data segments to obtain a reconstructed image sequence; performing multi-level consistency verification on the reconstructed image sequence, including pixel-level grayscale continuity verification, feature-level anatomical structure continuity verification, and semantic-level physiological parameter rationality verification, to generate complete localized image data; storing the complete localized image data in layers according to patient ID, examination item, and image sequence to obtain an execution result of the localized image acquisition and processing procedure; wherein the connected component analysis on the time-correlation graph to identify missing frames, damaged frames, and time-correlation abnormal regions in the image sequence, count the number of connected components and topological properties, and obtain a connected component analysis result includes: traversing all nodes of the time-correlation graph based on a depth-first search algorithm, identifying mutually connected node sets, obtaining connected components and the number of connected components; calculating topological property parameters for each connected component, including connectivity, diameter, clustering coefficient, and centrality index, to evaluate the integrity degree of the image sequence and obtain an integrity result; identifying sequence break positions caused by device failure, transmission interruption, or buffer overflow by analyzing the time span and spatial distribution characteristics of the connected components; verifying the rationality of the connected components based on physiological periodicity patterns, obtaining a physiological verification result for connected patterns corresponding to physiological periods in periodic image detection; constructing a connected component analysis result based on the number of connected components, the integrity result, the sequence break positions, and the physiological verification result; comparing the connected component analysis result with a pre-set integrity threshold, and marking any connected component with a connectivity degree lower than the pre-set threshold as an abnormal sequence that needs to be reconstructed to generate a sequence quality assessment report. An offline diagnosis processing module is configured to perform local AI algorithm analysis and lesion identification processing on ultrasonic image data collected in an offline state, generate an offline diagnosis report in combination with report template data in a local cache database, and complete an offline business closed loop; An incremental synchronization module is configured to continuously monitor the network connection state data, detect a network recovery event, and trigger a data synchronization process. The incremental synchronization module identifies to-be-synchronized data generated during the offline period by using a difference comparison algorithm, performs incremental data upload to a server, and completes offline collection and dynamic synchronization processing of multi-modal fusion ultrasonic images. The incremental synchronization module includes the following steps: constructing the to-be-synchronized data into a sparse graph structure, in which each data unit is a node, and the dependency relationship and conflict relationship between the data are connected as edges, and calculating a node weight value; applying a maximum weighted independent set algorithm based on the sparse graph structure to optimize a synchronization strategy, solving a data subset that can be synchronized in parallel and has the maximum weight by using a distributed approximate algorithm; processing the data subset in batches according to the dependency relationship, preferentially synchronizing emergency data and basic dependency data with high weight values, and avoiding data conflicts and transmission failures; in a multi-client environment, performing synchronization plan coordination by using a lightweight coordination protocol, preventing server load impact and optimizing network resource utilization efficiency; performing state write-back and local cleaning operations on the synchronized data, marking the synchronized data and clearing temporary storage, and obtaining an execution result of the incremental data upload to the server.

Images