Ultrasound image intelligent analysis system based on deep learning

Abstract

The present invention discloses an ultrasonic image intelligent analysis system based on deep learning, which includes a data cleaning module, a dynamic noise reduction module, a spatial normalization module, an image recognition model and a time series analysis model. The system operation process specifically includes the following steps: obtaining original ultrasonic image data; preprocessing the original ultrasonic image data to generate standardized image data; inputting the standardized image data into the image recognition model, extracting local texture features through the image recognition model, using the time series analysis model to capture dynamic motion features, combining spatial correlation analysis to establish global structural features; and performing pathological classification and diagnostic decision-making. The present invention has the following advantages and effects: the ultrasonic image intelligent analysis system based on deep learning improves the stability of image quality, the integrity of feature extraction and the depth of feature fusion, and ultimately significantly improves the quality of images and the accuracy of diagnosis, and has high clinical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic image intelligent analysis, and in particular to an ultrasonic image intelligent analysis system based on deep learning. Background Art

[0002] Ultrasound imaging, as a key diagnostic tool in medical imaging, is widely used in the examination and diagnosis of various clinical conditions. Its advantages, including non-invasiveness, real-time dynamic imaging, and low cost, have led to its widespread application in a variety of fields, including cardiovascular, oncology, and obstetrics and gynecology. However, ultrasound imaging is often affected by various factors (such as probe motion, patient positioning, and imaging equipment performance), resulting in noise, artifacts, and signal loss, which can affect diagnostic accuracy. Therefore, improving the quality of ultrasound images and enhancing their diagnostic accuracy has become a key research topic in the field of medical imaging.

[0003] In recent years, deep learning technology has made significant progress in image processing, particularly in medical image analysis. Through its powerful feature learning capabilities, deep learning can automatically extract useful information from images, effectively improving the automation and intelligence of image analysis. In ultrasound image processing, deep learning can automatically perform operations such as image recognition, noise reduction, restoration, feature extraction, and classification, achieving results comparable to or even exceeding those of traditional image processing methods.

[0004] However, the current ultrasound image analysis technology based on deep learning still has some technical bottlenecks and challenges. First, due to the particularity of its imaging principle, ultrasound images usually face problems such as excessive noise, large artifact areas, and signal loss. Existing noise reduction and restoration algorithms often cannot retain sufficient detail information while ensuring image quality, and therefore cannot provide sufficiently accurate diagnostic basis. Secondly, existing ultrasound image analysis methods often only consider single image features, such as local texture features or local dynamic features, and lack comprehensive consideration of global structural features. In order to further improve the accuracy and robustness of ultrasound image analysis, combining information from multiple features, especially local texture, dynamic motion, and global structural features, has become the key to improving the performance of intelligent analysis of ultrasound images. In addition, existing methods usually do not fully consider individual differences in cases and fail to automatically adjust the weight distribution of features according to different case types, which makes it impossible to effectively diagnose some special types of cases.

[0005] To address these issues, a growing number of studies are proposing comprehensive approaches that leverage a variety of image processing techniques and deep learning models to perform multi-level, multi-dimensional processing of ultrasound images. For example, a multi-stage collaborative processing pipeline architecture is used to pre-process images, including data cleaning, noise reduction, and standardization, to improve image quality from the source. Simultaneously, deep learning models are used to perform image recognition and time series analysis on images, extracting multiple features such as texture, dynamics, and structure. These features are then integrated through adaptive mechanisms to further enhance the accuracy and robustness of diagnostic results.

[0006] Furthermore, with the continuous development of deep learning technology, deep learning models based on convolutional neural networks, recurrent neural networks, and long-short-term memory networks are gaining recognition in ultrasound image analysis. These models can automatically learn features from large amounts of ultrasound image data, reducing manual intervention and detecting potential abnormalities in complex images. By combining temporal and spatial correlation analysis, deep learning models can capture the relationship between dynamic motion features and anatomical structural characteristics, providing a more accurate basis for pathological diagnosis.

[0007] In general, research on intelligent ultrasound image analysis technology is developing rapidly, and the application of deep learning technology has brought new opportunities for automated ultrasound image analysis. Although significant progress has been made in this field, problems such as unstable image quality, incomplete feature extraction, and insufficient feature fusion still need to be addressed. Summary of the Invention

[0008] The purpose of the present invention is to provide an ultrasonic image intelligent analysis system based on deep learning to solve the problems raised in the background technology.

[0009] The above technical objectives of the present invention are achieved through the following technical solutions:

[0010] The deep learning-based ultrasound imaging intelligent analysis system includes a data cleaning module, a dynamic noise reduction module, a spatial normalization module, an image recognition model, and a time series analysis model. The system operation process specifically includes the following steps:

[0011] S100, obtaining original ultrasound image data; wherein the original ultrasound image data includes an ultrasound scan sequence, each ultrasound scan sequence including grayscale images of multiple consecutive frames and corresponding spatiotemporal correlation information;

[0012] S200, preprocessing the raw ultrasound image data to generate standardized image data; wherein the preprocessing process is performed based on a multi-stage collaborative processing pipeline architecture, the pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial normalization module, and the modules achieve real-time interaction and state synchronization through a data bus;

[0013] S300: Input the standardized image data into the image recognition model, extract local texture features through the image recognition model, use the time series analysis model to capture dynamic motion features, combine spatial correlation analysis to establish global structural features, and use an adaptive weight allocation mechanism to fuse local texture features, dynamic motion features, and global structural features, and dynamically adjust according to the case type;

[0014] S400: Execute pathology classification and diagnosis decision based on the fused local texture features, dynamic motion features and global structural features.

[0015] By adopting the above technical solution, a multi-level collaborative processing pipeline architecture based on a data cleaning module, a dynamic noise reduction module, and a spatial normalization module can effectively remove noise and artifact areas in the image and repair signal loss areas. This not only improves the clarity of the image, but also ensures the structural consistency of the image and enhances the diagnostic value of the image. The modules interact and synchronize their status in real time through the data bus, ensuring the efficiency and seamless connection of the entire image processing process. The synergistic effect of data cleaning, noise reduction, and standardization enables ultrasound images to achieve high quality standards in the preprocessing stage, laying the foundation for subsequent image recognition and diagnostic decision-making. By extracting local texture features through an image recognition model, capturing dynamic motion features through a time series analysis model, and establishing global structural features through spatial correlation analysis, the system comprehensively considers multiple feature information. An adaptive weight allocation mechanism dynamically adjusts feature weights based on case type, further improving the accuracy of pathology classification and diagnostic decision-making and enhancing the system's adaptability to different case types. By integrating local, dynamic, and global feature information, the system comprehensively considers multidimensional features in the image during the diagnostic process, enabling more accurate identification of lesions and addressing complex image conditions such as noise, artifacts, and signal loss, effectively improving diagnostic accuracy and robustness. Furthermore, the system automatically adjusts feature weights based on the actual case type, ensuring that the feature analysis for each case is optimized based on its specific circumstances. This adaptive adjustment mechanism enables the system to make more accurate diagnoses despite individual differences, avoiding the misdiagnosis problem that often occurs in traditional methods due to insufficient standardization.

[0016] Further configuration is that the step S200 is specifically as follows:

[0017] S201. Perform the following operations in the data cleaning module:

[0018] Traversing all ultrasound scanning sequences, detecting whether there are artifact areas and signal loss areas in consecutive frames of each ultrasound scanning sequence;

[0019] If an artifact area exceeding a set threshold is detected in a frame of the ultrasound scan sequence, the adjacent frame compensation mechanism is activated and the median grayscale value of the same position area of the five previous and next frames in the ultrasound scan sequence is used to fill the gap.

[0020] If a signal loss area is detected in three or more consecutive frames of the ultrasound scanning sequence, the pixel matrix of the missing area is reconstructed using an elastic interpolation algorithm based on the probe motion trajectory data in the spatiotemporal correlation information;

[0021] S202: Execute multi-scale feature fusion denoising in the dynamic denoising module, specifically including:

[0022] Perform wavelet packet decomposition on the grayscale image of each ultrasound scanning sequence to generate a high-frequency sub-band coefficient matrix and a low-frequency sub-band coefficient matrix;

[0023] A dual-threshold noise reduction strategy based on adaptive signal-to-noise ratio adjustment is used for processing:

[0024] For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, the Bayesian shrinkage function is applied to suppress noise; when the local signal-to-noise ratio is higher than the set second threshold, the original coefficients are retained without processing;

[0025] Performing non-local means filtering on the low-frequency subband coefficient matrix;

[0026] S203, implementing multimodal registration in the spatial normalization module, specifically including:

[0027] An anatomical structure recognition model based on a deep neural network was constructed. The preprocessed ultrasound image data was input into the anatomical structure recognition model. The anatomical features of organ edges and vascular orientation were extracted through the convolutional layer, and the fully connected layer was used to classify and determine whether it was a standard section.

[0028] If it is a standard section, the subsequent processing is performed directly; if it is a non-standard section, the elastic registration process is triggered before the subsequent processing is performed;

[0029] Perform the following operations on the registered ultrasound image data:

[0030] Unify spatial resolution through bicubic interpolation;

[0031] Linearly map the original grayscale value to the 8-bit range and limit the overexposed area;

[0032] S204: Integrate the processing results of the data cleaning module, the dynamic noise reduction module, and the spatial normalization module, and output standardized image data that meets the spatial resolution, grayscale range, and anatomical consistency.

[0033] By adopting the above technical solutions, the data cleaning module can identify missing areas and artifacts in the image, and supplement them through interpolation and repair techniques of similar image areas to ensure the integrity of the image; the dynamic noise reduction module adopts a noise reduction algorithm based on convolutional neural networks, which can effectively remove noise in the image while retaining detail information, avoiding the loss of valuable structural information during the denoising process; the spatial standardization module uses an automated alignment method to ensure that image data collected at different times and with different devices have consistent spatial standards, thereby achieving the unification of multi-source image data; the above process greatly improves image quality and maintains the integrity of important detail information in the image. The optimized denoising effect makes subtle lesions more obvious, especially in ultrasound images. Since the influence of noise often obscures the lesion area, the addition of this module effectively improves the clarity and recognizability of the image.

[0034] Further configuration is that, in the step S201, the elastic interpolation algorithm is used to reconstruct the pixel matrix of the missing area specifically as follows:

[0035] When a signal loss area is detected for three or more consecutive frames in the ultrasound scanning sequence, the loss time window is locked and the discrete coordinate data set of the corresponding time period is extracted;

[0036] The discrete coordinate data set is fitted into a continuous motion trajectory curve through B-spline curve, eliminating the coordinate mutation caused by operation jitter and generating a high-precision probe motion path;

[0037] Select anatomical feature points as control points in normal frames before and after signal loss, and establish a spatiotemporal mapping table of time-probe coordinates-anatomical point coordinates;

[0038] Based on the spatiotemporal mapping table, the ultrasound scanning area is modeled as an elastic thin plate model, the regional stiffness parameters are set, and the anatomical control points are bound as thin plate anchor points;

[0039] Coarse-grained structure restoration and fine-grained detail optimization are performed, and the restored image sequence is generated by combining elastic thin plate model constraints with medical rule base verification.

[0040] By adopting the above technical solutions, traditional image restoration methods often rely on simple linear interpolation or neighboring interpolation methods. Although these methods are simple, they often lead to obvious artifacts or distortion in the repaired area for complex ultrasound images. The elastic interpolation algorithm, on the other hand, uses adaptive curve fitting technology to accurately reconstruct the missing area and naturally blend it with the structure and texture of the surrounding area, avoiding information loss or over-smoothing in image restoration, and the restored image has high visual continuity and consistency. Especially when dealing with missing areas caused by equipment vibration or changes in patient position, the elastic interpolation algorithm shows high restoration quality, which can effectively restore important structural information in the image and improve the integrity and credibility of the image.

[0041] Further configuration is that, in step S203, the elastic registration process is triggered specifically as follows:

[0042] Based on the preprocessed ultrasound image data and the associated spatiotemporal information, a multi-resolution B-spline deformation field is constructed;

[0043] Match the overall shape at low resolution and adjust the local details at high resolution, and adjust the regularization parameters.

[0044] By adopting the above technical solution, spatiotemporal registration technology can accurately align each frame of the image in the time series. Especially when the patient moves or the scanning angle changes, it can intelligently judge the differences between different frames through the combination of spatiotemporal information and use the elastic deformation field for precise registration. This registration method not only considers the spatial dimension of the image, but also makes full use of the information of the temporal dimension to ensure the global consistency of the image. At the same time, during the registration process, the alignment between images is gradually optimized through a multi-resolution method, ensuring accurate registration effect at high resolution, thereby improving the detail retention of the image and avoiding structural distortion caused by registration errors.

[0045] Further configuration is that, in the step S300, local texture features are extracted by the image recognition model, dynamic motion features are captured by the time series analysis model, and global structural features are established in combination with spatial correlation analysis. Specifically:

[0046] Use pre-trained image recognition models to analyze standardized image data and automatically identify local texture features, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity;

[0047] Through the time series analysis model, the continuous frame sequence in the standardized image data is dynamically tracked to quantify the dynamic motion characteristics;

[0048] The single-frame images in the standardized imaging data are divided into grids, and the organ position relationship and vascular network topology are modeled through spatial correlation analysis to generate global structural features that represent the anatomical structure.

[0049] By adopting the above technical solutions, multi-level feature information can be extracted from images. The image recognition model focuses on capturing local texture features, accurately identifying subtle lesion markers such as the edges of tumors and changes in blood flow. The time series analysis model captures dynamic features during motion, such as organ movement and tumor changes, through time series data analysis. Spatial correlation analysis combines global structural information to identify possible lesion areas within a large area. The comprehensive extraction of these features enables the system to comprehensively analyze ultrasound images and identify more detailed lesion areas. The combination of local features and global structure makes the system more accurate in diagnosing lesions, avoiding missed diagnoses caused by single feature analysis.

[0050] Further configuration is that, in the step S300, an adaptive weight allocation mechanism is used to fuse local texture features, dynamic motion features and global structural features, and the feature weight ratio is dynamically adjusted according to the case type, specifically:

[0051] The local texture features, dynamic motion features and global structural features are standardized to eliminate the differences in numerical dimensions.

[0052] Configure the initial weights of local texture features, dynamic motion features, and global structure features;

[0053] Dynamically adjust the weight ratio according to the actual case type:

[0054] Temperature scaling technology is used to calibrate the probability of fused features and optimize classification confidence.

[0055] By adopting the above technical solution, local texture features, dynamic motion features, and global structural features each carry different information in the image data. For example, local texture features describe the microstructure in the image, dynamic motion features reflect the dynamic changes of organs and lesion areas, and global structural features provide a macroscopic view of the entire anatomical structure. The introduction of an adaptive weight allocation mechanism can dynamically adjust the weight ratio of these features according to the actual needs of different case types. This adjustment mechanism enables the system to make the most appropriate analysis for different types of cases, which not only improves diagnostic accuracy but also reduces over-reliance on specific features and avoids misdiagnosis or missed diagnosis due to a single feature. By standardizing the three types of features, any possible differences in numerical dimensions between them are eliminated, ensuring that they have a relatively balanced influence during feature fusion. After feature fusion, the fused features are probabilistically calibrated using temperature scaling technology to further optimize the classification confidence.

[0056] Further configuration is that the step S400 is specifically as follows:

[0057] S401, receiving the feature data fused in step S300, and performing multi-task diagnosis simultaneously:

[0058] Automatically detect suspicious lesion areas in the standardized image data, and verify the rationality of the anatomical position in combination with the standard section data generated in step S203;

[0059] Analyze dynamic motion characteristics and generate motion trajectory graphs that change over time;

[0060] Establish an organ position relationship model through spatial grid division to identify abnormal blood vessel orientation or tissue deformation;

[0061] S402, hierarchical verification and correction;

[0062] S403, generating a visual diagnosis report;

[0063] S404, establishing a self-optimizing system;

[0064] S405: Output the final structured report.

[0065] By adopting the above technical solution, in step S401, the system not only automatically detects suspicious lesion areas in the standardized imaging data, but also verifies the rationality of the anatomical position in combination with the standard section data. This capability enables the system to determine whether there is a lesion area through more accurate anatomical structure information, and can provide immediate feedback in the actual clinical environment, helping doctors to quickly confirm the lesion site and conduct further examinations; the analysis of dynamic motion characteristics and the organ position relationship model established by spatial grid division enable the system to have stronger diagnostic capabilities in tracking tumor growth, organ deformation, and blood vessel direction.

[0066] Further configuration is that the step S402 is specifically as follows:

[0067] Compare the preliminary results with the preset medical standards; if any discrepancies are found, the multimodal registration process in step S203 is automatically triggered to recalibrate the data; for cases with questionable diagnosis, return to step S300 to adjust the weight distribution mechanism and conduct a secondary analysis;

[0068] Check whether the diagnostic conclusions of 5 consecutive frames of ultrasound image data are consistent, and call the adjacent frame compensation mechanism of step S201 to repair the data for the sudden change result.

[0069] By adopting the above technical solution, in step S402, when the preliminary diagnosis result conflicts with the preset medical standard, the system can automatically trigger the multimodal registration process for calibration. This process effectively avoids diagnostic errors caused by image deviation or error, ensuring the accuracy of the ultrasound image data. By examining five consecutive frames of image data, the system can confirm the consistency of the diagnostic conclusion, thereby avoiding the risk of single-frame misdiagnosis. For questionable cases, the system will return to step S300 according to the scheduling mechanism to adjust the weight distribution mechanism and re-analyze. This step provides a second chance for diagnosis, ensuring that the system can provide more accurate analysis results when facing complex lesions. Through this mechanism, the system's error rate is significantly reduced, further enhancing its reliability in clinical practice. By comparing and analyzing five consecutive frames of ultrasound image data, the system can confirm the consistency of the diagnostic results and compensate for sudden changes. This mechanism enhances the model's adaptability to dynamic lesions or instantaneous changes, ensuring real-time tracking and accurate diagnosis of patients' conditions.

[0070] Further configuration is that the step S403 is specifically as follows:

[0071] Convert the dynamic motion features extracted in step S300 into a motion curve graph, and combine it with the noise-reduced data in step S202 to generate a clear spectrum graph;

[0072] Semi-transparent color blocks are superimposed on the standardized image data to mark the lesion area, and the HSV color gamut mapping technology is used to convert the diagnostic confidence into hue saturation parameters;

[0073] Automatically match diagnostic templates based on analysis results and highlight key judgment criteria.

[0074] By adopting the above technical solution, by converting dynamic motion features into motion curve graphs and combining them with the noise-reduced data to generate clear spectrum graphs, the system can intuitively present the changing trends and dynamic features of the lesion area. Doctors can quickly judge the type, progression and relationship of the lesion with other organs by observing these graphs, thereby accelerating diagnostic decisions; superimposing semi-transparent color blocks on the standardized image data to mark the lesion area, and using HSV color gamut mapping technology to convert diagnostic confidence into hue saturation parameters. This color block marking method not only makes the lesion area more prominent, but also can intuitively reflect the system's confidence in the diagnostic results, helping doctors make decisions more clearly and quickly.

[0075] Further configuration is that the step S404 is specifically as follows:

[0076] Collect the difference data between the clinical diagnosis results and the multi-task diagnosis in step S401, and synchronously update the parameters of the image recognition model, the time series analysis model, and the adaptive weight allocation mechanism through the back propagation algorithm;

[0077] When the misjudgment rate of a certain type of case exceeds a set threshold, the initial weight ratios of the local texture features, dynamic motion features, and global structural features in step S300 are automatically optimized;

[0078] The optimized parameters are regularly fed back to steps S201 to S203 to improve the accuracy of artifact restoration and multimodal image registration.

[0079] By adopting the above technical solution, the system can collect the difference data between clinical diagnosis results and multi-task diagnosis through the backpropagation algorithm, and synchronously update the parameters of the image recognition model, time series analysis model and adaptive weight allocation mechanism. This process enables the system to automatically adjust and optimize its algorithm according to new clinical data and diagnostic conditions, ensuring that it still maintains a high diagnostic accuracy when facing new cases; when the misjudgment rate of a certain type of case exceeds the set threshold, the system will automatically optimize the feature weight ratio in step S300, thereby improving the performance of the model in specific types of cases. This mechanism ensures that the system can always maintain good performance while continuously receiving new data, and the diagnostic ability of the system is continuously enhanced over time; the subsequent feedback mechanism ensures that ultrasound images can still provide high-quality data input in various complex situations, avoiding misdiagnosis and missed diagnosis due to poor image quality.

[0080] In summary, the present invention has the following beneficial effects:

[0081] The deep learning-based intelligent ultrasound image analysis system significantly improves the image quality and diagnostic accuracy through multi-level and multi-dimensional image processing and analysis technology, and has high clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Schematic diagram of the process of the embodiment. DETAILED DESCRIPTION

[0083] The present invention will be further described in detail below with reference to the accompanying drawings.

[0084] As attached Figure 1 As shown;

[0085] This embodiment discloses an ultrasonic image intelligent analysis system based on deep learning, which includes a data cleaning module, a dynamic noise reduction module, a spatial normalization module, an image recognition model, and a time series analysis model. The system operation process specifically includes the following steps:

[0086] S100, obtaining original ultrasound image data; wherein the original ultrasound image data includes an ultrasound scan sequence, each ultrasound scan sequence including grayscale images of multiple consecutive frames and corresponding spatiotemporal correlation information;

[0087] S200, preprocessing the raw ultrasound image data to generate standardized image data; wherein the preprocessing process is performed based on a multi-stage collaborative processing pipeline architecture, the pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial normalization module, and the modules achieve real-time interaction and state synchronization through a data bus;

[0088] S300: Input the standardized image data into the image recognition model, extract local texture features through the image recognition model, use the time series analysis model to capture dynamic motion features, combine spatial correlation analysis to establish global structural features, and use an adaptive weight allocation mechanism to fuse local texture features, dynamic motion features, and global structural features, and dynamically adjust according to the case type;

[0089] S400: Execute pathology classification and diagnosis decision based on the fused local texture features, dynamic motion features and global structural features.

[0090] Herein, step S200 is specifically as follows:

[0091] S201. Perform the following operations in the data cleaning module:

[0092] Traversing all ultrasound scanning sequences, detecting whether there are artifact areas and signal loss areas in consecutive frames of each ultrasound scanning sequence;

[0093] If an artifact area is detected in a frame of the ultrasound scan sequence and the area ratio exceeds the set threshold (the threshold is set to 15%), the adjacent frame compensation mechanism is activated and the median grayscale value of the same position area in the five frames before and after the ultrasound scan sequence is used to fill the gap.

[0094] If a signal loss area is detected in three or more consecutive frames of the ultrasound scanning sequence, the pixel matrix of the missing area is reconstructed using an elastic interpolation algorithm based on the probe motion trajectory data in the spatiotemporal correlation information;

[0095] S202: Execute multi-scale feature fusion denoising in the dynamic denoising module, specifically including:

[0096] Perform wavelet packet decomposition on the grayscale image of each ultrasound scanning sequence to generate a high-frequency sub-band coefficient matrix and a low-frequency sub-band coefficient matrix;

[0097] A dual-threshold noise reduction strategy based on adaptive signal-to-noise ratio adjustment is used for processing:

[0098] For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, the Bayesian shrinkage function is applied to perform noise suppression; when the local signal-to-noise ratio is higher than the set second threshold, the original coefficients are retained without processing; the set first threshold is 8dB, and the set second threshold is 15dB;

[0099] Perform non-local mean filtering on the low-frequency subband coefficient matrix, specifically setting the search window radius to 7 pixels, the similarity window radius to 3 pixels, and the Gaussian weighting parameter to 0.6;

[0100] S203, implementing multimodal registration in the spatial normalization module, specifically including:

[0101] An anatomical structure recognition model based on a deep neural network was constructed. The preprocessed ultrasound image data was input into the anatomical structure recognition model. The anatomical features of organ edges and vascular orientation were extracted through the convolutional layer, and the fully connected layer was used to classify and determine whether it was a standard section.

[0102] If it is a standard section, the subsequent processing is performed directly; if it is a non-standard section, the elastic registration process is triggered before the subsequent processing is performed;

[0103] Perform the following operations on the registered ultrasound image data:

[0104] The spatial resolution is unified to 0.2 mm × 0.2 mm through bicubic interpolation;

[0105] Linearly map the original grayscale value to the 8-bit range and limit the overexposed area;

[0106] S204: Integrate the processing results of the data cleaning module, the dynamic noise reduction module, and the spatial normalization module, and output standardized image data that meets the spatial resolution, grayscale range, and anatomical consistency.

[0107] In step S201, the elastic interpolation algorithm is used to reconstruct the missing area pixel matrix specifically as follows:

[0108] When a signal loss region is detected in three or more consecutive frames of the ultrasound scanning sequence, the loss time window is locked and the discrete coordinate data set of the corresponding time period is extracted. The criterion for determining the signal loss region is that the standard deviation of the grayscale value of the same anatomical region in three consecutive frames of the ultrasound scanning sequence is less than 5% of the normal value.

[0109] The discrete coordinate data set is fitted into a continuous motion trajectory curve through B-spline curve, eliminating the coordinate mutation caused by operation jitter and generating a high-precision probe motion path;

[0110] Among them, the B-spline curve fitting is specifically:

[0111] Perform cubic B-spline interpolation on discrete coordinates to generate a continuous trajectory curve;

[0112] When the distance between adjacent coordinate points suddenly changes by more than 2 mm, the trajectory smoothing correction algorithm is activated to remove abnormal points and refit.

[0113] Select anatomical feature points as control points in normal frames before and after signal loss, and establish a spatiotemporal mapping table of time-probe coordinates-anatomical point coordinates;

[0114] Among them, the selection of anatomical feature points includes:

[0115] Static control points: bone markers, vascular bifurcation center points, with a coordinate change rate of ≤0.1 mm / frame;

[0116] Dynamic control point: the thinnest point of the myocardial wall and the center point of valve opening and closing. The angle between its movement direction and the tangent direction of the probe trajectory is ≤15°.

[0117] Based on the spatiotemporal mapping table, the ultrasound scanning area is modeled as an elastic thin plate model, the regional stiffness parameters are set, and the anatomical control points are bound as thin plate anchor points;

[0118] The parameters of the elastic thin plate model are set so that the stiffness coefficient of the myocardial region is 10 times that of the blood flow region. The thin plate deformation constraints include: myocardial stretching limit ≤ 0.2 mm / frame, and vascular branch angle change ≤ 5° / frame.

[0119] Coarse-grained structure restoration and fine-grained detail optimization are performed, combined with elastic thin plate model constraints and medical rule base verification to generate the restored image sequence. The coarse-grained structure restoration process involves projecting the anatomical structure of five adjacent frames along a trajectory onto the lost frame based on the elastic thin plate model to generate an initial grayscale distribution. The fine-grained detail optimization process involves gradient sharpening and adding motion blur to tissue boundaries under the thin plate stiffness constraints.

[0120] In step S203, the elastic registration process is triggered as follows:

[0121] Based on the preprocessed ultrasound image data and the associated spatiotemporal information, a multi-resolution B-spline deformation field is constructed;

[0122] At low resolution, the overall shape is matched, the deformation field grid spacing is 16 mm, and the regularization parameter is set to 0.3. The purpose is to roughly align the non-standard section to the standard template. At high resolution, local details are adjusted, and the regularization parameter is adjusted. The deformation field grid spacing is reduced to 4 mm, and the value of the regularization parameter will be reduced. The purpose is to correct details such as myocardial wall thickness and vascular bifurcation angle.

[0123] In step S300, local texture features are extracted through the image recognition model, dynamic motion features are captured using the time series analysis model, and global structural features are established by combining spatial correlation analysis. Specifically:

[0124] Using a pre-trained image recognition model and the ResNet-50 network, standardized image data is analyzed to automatically identify local texture features, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity. Specifically:

[0125] A hierarchical and progressive feature extraction strategy is used to capture basic texture patterns through low-level convolutions of the ResNet-50 network, and high-level convolutions are used to identify complex pathological patterns.

[0126] Multi-scale feature fusion is implemented, and feature maps at different levels are upsampled to a unified resolution and then channel-wise spliced. Finally, by embedding a channel attention mechanism, the noise channel is automatically suppressed and the response weights of lesion-related channels are enhanced.

[0127] Through the time series analysis model, the LSTM network is used to dynamically track the continuous frame sequence in the standardized image data, and quantify the dynamic motion characteristics such as myocardial motion rate and valve opening and closing cycle;

[0128] The single-frame image in the standardized imaging data is gridded into 16×16 grid blocks. The positional relationship of organs and the topological structure of vascular networks are modeled through spatial correlation analysis to generate global structural features representing the anatomical structure. Specifically:

[0129] Perform 16×16 grid division on the standardized single-frame image to generate 256 local blocks;

[0130] Add a spatial position code to each block and record its coordinate offset relative to the anatomical landmark;

[0131] The multi-head self-attention mechanism is used to calculate the correlation weights between blocks, focusing on strengthening the following medically significant areas:

[0132] Matching relationship between symmetrical regions of the ventricle;

[0133] Connectivity pathways at vascular bifurcations;

[0134] Continuity characteristics of organizational boundaries.

[0135] In step S300, an adaptive weight allocation mechanism is used to fuse local texture features, dynamic motion features, and global structural features, and the feature weight ratio is dynamically adjusted according to the case type. Specifically,

[0136] The local texture features, dynamic motion features and global structural features are standardized to eliminate the differences in numerical dimensions.

[0137] Configure the initial weights of local texture features, dynamic motion features, and global structure features;

[0138] The initial weights are set to:

[0139] Local texture feature weight 40%;

[0140] Dynamic motion feature weight 30%;

[0141] The weight of global structural features is 30%;

[0142] Dynamically adjust the weight ratio according to the actual case type:

[0143] For example, when dealing with static lesions such as calcified plaques, the weight of local texture features is increased to 60%;

[0144] When analyzing dynamic abnormalities such as heart valve regurgitation, the weight of dynamic motion features is increased to 50%;

[0145] Temperature scaling technology is used to calibrate the probability of fused features and optimize classification confidence.

[0146] Herein, step S400 is specifically as follows:

[0147] S401, receiving the feature data fused in step S300, and performing multi-task diagnosis simultaneously:

[0148] Automatically detect suspicious lesion areas in the standardized image data, and verify the rationality of the anatomical position in combination with the standard section data generated in step S203;

[0149] Analyze dynamic motion characteristics and generate motion trajectory graphs that change over time;

[0150] Establish an organ position relationship model through spatial grid division to identify abnormal blood vessel orientation or tissue deformation;

[0151] S402, hierarchical verification and correction;

[0152] S403, generating a visual diagnosis report;

[0153] S404, establishing a self-optimizing system;

[0154] S405: Output the final structured report.

[0155] Among them, step S402 is specifically as follows:

[0156] The preliminary results are compared with pre-set medical standards, such as the normal range of ventricular contraction velocity. If any discrepancies are found, the multimodal registration process in step S203 is automatically triggered to recalibrate the data. For cases with questionable diagnoses, the process returns to step S300 to adjust the weight distribution mechanism, such as increasing the local texture weight from 40% to 60%, for secondary analysis.

[0157] Check whether the diagnostic conclusions of 5 consecutive frames of ultrasound image data are consistent, and call the adjacent frame compensation mechanism of step S201 to repair the data for the sudden change result.

[0158] Among them, step S403 is specifically as follows:

[0159] Convert the dynamic motion features extracted in step S300 into a motion curve graph, and combine it with the noise-reduced data in step S202 to generate a clear spectrum graph;

[0160] A semi-transparent color block was superimposed on the standardized image data to mark the lesion area, and the HSV color gamut mapping technology was used to convert the diagnostic confidence into hue saturation parameters, where red indicates a confidence level greater than 85%, yellow indicates 50%-85%, and blue indicates less than 50%.

[0161] Automatically match diagnostic templates based on analysis results and highlight key judgment criteria.

[0162] Among them, step S404 is specifically as follows:

[0163] Collect the difference data between the clinical diagnosis results and the multi-task diagnosis in step S401, and synchronously update the following parameters through the backpropagation algorithm: the convolution kernel weights of the image recognition model, the LSTM gate unit parameters of the time series analysis model, and the feature scale coefficient in the adaptive weight allocation mechanism;

[0164] When the misclassification rate of a certain type of case exceeds the set threshold of 5%, the initial weight ratios of local texture features, dynamic motion features, and global structural features in step S300 are automatically optimized, and the noise reduction intensity in step S202 is adjusted in conjunction;

[0165] The optimized parameters are regularly fed back to steps S201 to S203 to improve the accuracy of artifact restoration and multimodal image registration.

[0166] Application Example 1

[0167] In step S100, ultrasound imaging data of 200 patients with suspected mitral regurgitation were obtained from the medical records of a tertiary hospital. Each case included a 30-second continuous ultrasound scan sequence, with a total data volume of 180,000 frames. Artifacts were detected in 12% of the original ultrasound image frames, with an average area of 18%. 8% of the ultrasound scan sequences had 3-5 consecutive frames of signal loss, and the standard deviation of the grayscale in the myocardial area decreased from the normal value of 45±8 to 2±1.

[0168] In step S201, the adjacent frame compensation mechanism is activated, and the artifact area is filled with the median of the adjacent five frames. The image structure similarity index is improved from 0.72 to 0.89.

[0169] When the elastic interpolation algorithm reconstructs the signal-loss area, a cubic B-spline curve is used to fit the probe trajectory. Combined with an elastic thin plate model, the elastic thin plate model is configured with a myocardial stiffness coefficient of 120 kPa and a blood flow area of 12 kPa. The error between the reconstructed valve motion trajectory and the gold standard MRI is 0.8 mm, which is less than the clinically allowable error of 1.5 mm.

[0170] In step S202, the db4 basis function is used for wavelet packet decomposition, and the high-frequency sub-band Bayesian shrinkage is performed to improve the contrast-to-noise ratio of the myocardial boundary from 12 dB to 24 dB;

[0171] Low-frequency non-local mean filtering improves tissue texture clarity by 37% and optimizes computational time to 18ms / frame.

[0172] In step S203, the anatomical structure recognition model accurately identifies the standard four-chamber heart view and performs multi-resolution registration on the non-standard view, with 16mm grid coarse registration and 4mm grid fine registration. After registration, the ventricular volume measurement error is reduced from 9.2ml to 2.1ml.

[0173] In step S300, the local texture features extracted by the image recognition model include 32 layers of convolutional feature maps, and the response value in the valve calcification area is increased by 3.6 times.

[0174] The time series analysis model accurately captures the change in the velocity of the regurgitant beam, with the peak velocity changing from 2.8 m / s to 4.2 m / s;

[0175] The spatial meshing modeled ventricular symmetry, with a sensitivity of 92% for detecting abnormal areas of left atrial / left ventricular volume ratio in regurgitation cases;

[0176] Adaptive weight allocation increases the weight of dynamic motion features to 50%, and the area under the curve of the fusion model on the test set reaches 0.96, which is significantly better than the area under the curve of the single feature model between 0.82 and 0.89.

[0177] In step S400, the hierarchical verification found 5 false negatives, triggering a weight adjustment. That is, after the local texture weight was increased from 40% to 55%, the detection rate increased to 98.5%;

[0178] The self-optimization system has updated model parameters 12,000 times, reducing the myocardial motion trajectory prediction error by 42%;

[0179] The final diagnostic accuracy rate was 91.3%, and the diagnosis time was shortened from the traditional 45 minutes to 8 minutes.

[0180] Application Example 2

[0181] In step S100, ultrasound data of 350 patients with chronic liver disease were obtained from the case database of a tertiary hospital. Each case included a dynamic sequence of the portal vein long axis. 18% of the original ultrasound images contained rib artifacts, accounting for 23% of the area, and the continuous frame signal loss rate at the portal vein branches was 9%.

[0182] In step S201, when reconstructing the portal vein branches using the elastic interpolation algorithm, the vascular stiffness coefficient is set to 80 kPa, and the reconstructed vascular diameter error is 0.3 mm, which is less than the clinical standard of 0.5 mm;

[0183] In step S202, high-frequency sub-band double-threshold noise reduction improves the contrast-to-noise ratio of the liver lobule structure from 18dB to 31dB; low-frequency non-local means filtering preserves the portal sheath structure and improves the texture uniformity index by 29%;

[0184] In step S203, after registration, the standard deviation of the right hepatic vein angle is reduced from 7.2° to 1.8°, and the spatial resolution is unified to 0.2 mm / px, so that the fiber interval measurement accuracy reaches 0.05 mm;

[0185] In step S300, the image recognition model identifies lattice-like changes in the liver parenchyma, with a specificity of 91% for fibrosis above F3;

[0186] The time series analysis model captured portal vein pulsatility attenuation, with the pulsatility index changing from 0.38 to 0.22, and the predicted hepatic venous pressure gradient error was 3.2 mmHg;

[0187] Adaptive weight allocation focused on local texture features, increasing to 60%. The fusion model's fibrosis staging accuracy, with an overall Kappa value of 0.89, was significantly better than ultrasound elastography.

[0188] In step S400, the multi-task detection found 32 cases of portal vein thrombosis with a sensitivity of 95%, and the thrombus range was accurately located through HSV color gamut mapping;

[0189] Compared with the gold standard of liver biopsy, the diagnostic consistency rate was 89.4%, shortening the diagnosis time from 3 days to real-time analysis.

[0190] The above two examples verify the excellent performance of this method in dynamic cardiac diagnosis and static liver lesion analysis. All indicators exceed clinical requirements, proving its effectiveness.

[0191] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. Ultrasound image intelligent analysis system based on deep learning, characterized by: It includes a data cleaning module, a dynamic noise reduction module, a spatial normalization module, an image recognition model, and a time series analysis model. The system operation process specifically includes the following steps: S100, obtaining original ultrasound image data; wherein the original ultrasound image data includes an ultrasound scan sequence, each ultrasound scan sequence including grayscale images of multiple consecutive frames and corresponding spatiotemporal correlation information; S200, preprocessing the raw ultrasound image data to generate standardized image data; wherein the preprocessing process is performed based on a multi-stage collaborative processing pipeline architecture, the pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial normalization module, and the modules achieve real-time interaction and state synchronization through a data bus; S300: Input standardized image data into an image recognition model, extract local texture features through the image recognition model, capture dynamic motion features using a time series analysis model, establish global structural features through spatial correlation analysis, and use an adaptive weight allocation mechanism to fuse local texture features, dynamic motion features, and global structural features, and dynamically adjust based on case type; including: Use pre-trained image recognition models to analyze standardized image data and automatically identify local texture features, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity; Through the time series analysis model, the continuous frame sequence in the standardized image data is dynamically tracked to quantify the dynamic motion characteristics; The single-frame images in the standardized imaging data are divided into grids, and the positional relationship of organs and the topological structure of vascular networks are modeled through spatial correlation analysis to generate global structural features representing the anatomical structure. The local texture features, dynamic motion features and global structural features are standardized to eliminate the differences in numerical dimensions. Configure the initial weights of local texture features, dynamic motion features, and global structure features; Dynamically adjust the weight ratio according to the actual case type: Temperature scaling technology is used to calibrate the probability of fusion features and optimize classification confidence; S400: Execute pathology classification and diagnosis decision based on the fused local texture features, dynamic motion features and global structural features.

2. The deep learning-based ultrasonic image intelligent analysis system according to claim 1, characterized in that: The step S200 is specifically as follows: S201. Perform the following operations in the data cleaning module: Traversing all ultrasound scanning sequences, detecting whether there are artifact areas and signal loss areas in consecutive frames of each ultrasound scanning sequence; If an artifact area exceeding a set threshold is detected in a frame of the ultrasound scan sequence, the adjacent frame compensation mechanism is activated and the median grayscale value of the same position area of the five previous and next frames in the ultrasound scan sequence is used to fill the gap. If a signal loss area is detected in three or more consecutive frames of the ultrasound scanning sequence, the pixel matrix of the missing area is reconstructed using an elastic interpolation algorithm based on the probe motion trajectory data in the spatiotemporal correlation information; S202: Execute multi-scale feature fusion denoising in the dynamic denoising module, specifically including: Perform wavelet packet decomposition on the grayscale image of each ultrasound scanning sequence to generate a high-frequency sub-band coefficient matrix and a low-frequency sub-band coefficient matrix; A dual-threshold noise reduction strategy based on adaptive signal-to-noise ratio adjustment is used for processing: For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, the Bayesian shrinkage function is applied to suppress noise; when the local signal-to-noise ratio is higher than the set second threshold, the original coefficients are retained without processing; Performing non-local means filtering on the low-frequency subband coefficient matrix; S203, implementing multimodal registration in the spatial normalization module, specifically including: An anatomical structure recognition model based on a deep neural network was constructed. The preprocessed ultrasound image data was input into the anatomical structure recognition model. The anatomical features of organ edges and vascular orientation were extracted through the convolutional layer, and the fully connected layer was used to classify and determine whether it was a standard section. If it is a standard section, the subsequent processing is performed directly; if it is a non-standard section, the elastic registration process is triggered before the subsequent processing is performed; Perform the following operations on the registered ultrasound image data: Unify spatial resolution through bicubic interpolation; Linearly map the original grayscale value to the 8-bit range and limit the overexposed area; S204: Integrate the processing results of the data cleaning module, the dynamic noise reduction module, and the spatial normalization module, and output standardized image data that meets the spatial resolution, grayscale range, and anatomical consistency.

3. The deep learning-based ultrasonic image intelligent analysis system according to claim 2, characterized in that: In step S201, the elastic interpolation algorithm is used to reconstruct the missing area pixel matrix specifically as follows: When a signal loss area is detected for three or more consecutive frames in the ultrasound scanning sequence, the loss time window is locked and the discrete coordinate data set of the corresponding time period is extracted; The discrete coordinate data set is fitted into a continuous motion trajectory curve through B-spline curve, eliminating the coordinate mutation caused by operation jitter and generating a high-precision probe motion path; Select anatomical feature points as control points in normal frames before and after signal loss, and establish a spatiotemporal mapping table of time-probe coordinates-anatomical point coordinates; Based on the spatiotemporal mapping table, the ultrasound scanning area is modeled as an elastic thin plate model, the regional stiffness parameters are set, and the anatomical control points are bound as thin plate anchor points; Coarse-grained structure restoration and fine-grained detail optimization are performed, and the restored image sequence is generated by combining elastic thin plate model constraints with medical rule base verification.

4. The deep learning-based ultrasonic image intelligent analysis system according to claim 3, characterized in that: In step S203, the elastic registration process is triggered as follows: Based on the preprocessed ultrasound image data and the associated spatiotemporal information, a multi-resolution B-spline deformation field is constructed; Match the overall shape at low resolution and adjust the local details at high resolution, and adjust the regularization parameters.

5. The deep learning-based ultrasonic image intelligent analysis system according to claim 1, characterized in that: The step S400 is specifically as follows: S401, receiving the feature data fused in step S300, and performing multi-task diagnosis simultaneously: Automatically detect suspicious lesion areas in the standardized image data, and verify the rationality of the anatomical position in combination with the standard section data generated in step S203; Analyze dynamic motion characteristics and generate motion trajectory graphs that change over time; Establish an organ position relationship model through spatial grid division to identify abnormal blood vessel orientation or tissue deformation; S402, hierarchical verification and correction; S403, generating a visual diagnosis report; S404, establishing a self-optimizing system; S405: Output the final structured report.

6. The deep learning-based ultrasonic image intelligent analysis system according to claim 5, characterized in that: The step S402 is specifically as follows: Compare the preliminary results with the preset medical standards; if any discrepancies are found, the multimodal registration process in step S203 is automatically triggered to recalibrate the data; for cases with questionable diagnosis, return to step S300 to adjust the weight distribution mechanism and conduct a secondary analysis; Check whether the diagnostic conclusions of 5 consecutive frames of ultrasound image data are consistent, and call the adjacent frame compensation mechanism of step S201 to repair the data for the sudden change result.

7. The deep learning-based ultrasonic image intelligent analysis system according to claim 6, characterized in that: The step S403 is specifically as follows: Convert the dynamic motion features extracted in step S300 into a motion curve graph, and combine it with the noise-reduced data in step S202 to generate a clear spectrum graph; Semi-transparent color blocks are superimposed on the standardized image data to mark the lesion area, and the HSV color gamut mapping technology is used to convert the diagnostic confidence into hue saturation parameters; Automatically match diagnostic templates based on analysis results and highlight key judgment criteria.

8. The deep learning-based ultrasonic image intelligent analysis system according to claim 7, characterized in that: The step S404 is specifically as follows: Collect the difference data between the clinical diagnosis results and the multi-task diagnosis in step S401, and synchronously update the parameters of the image recognition model, the time series analysis model, and the adaptive weight allocation mechanism through the back propagation algorithm; When the misjudgment rate of a certain type of case exceeds a set threshold, the initial weight ratios of the local texture features, dynamic motion features, and global structural features in step S300 are automatically optimized; The optimized parameters are regularly fed back to steps S201 to S203 to improve the accuracy of artifact restoration and multimodal image registration.

Images