Ultrasonic image intelligent analysis system based on deep learning

Abstract

The invention discloses an ultrasonic image intelligent analysis system based on deep learning, which comprises a data cleaning module, a dynamic noise reduction module, a space standardization module, an image recognition model and a time sequence analysis model, and the system operation process specifically comprises 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 an image recognition model, extracting local texture features through the image recognition model, capturing dynamic motion features by using a time sequence analysis model, and establishing global structure features in combination with spatial correlation analysis; and executing pathological classification and diagnosis decisions. Compared with the prior art, the ultrasonic image intelligent analysis system based on deep learning has the following advantages and effects that the image quality stability, the feature extraction integrity and the feature fusion depth are improved, the image quality and the diagnosis accuracy are remarkably improved finally, and the ultrasonic image intelligent analysis system based on deep learning has very high clinical application value.

Description

Technical Field

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

[0002] As an important diagnostic tool in medical imaging, ultrasonic images are widely used in the examination and diagnosis of various clinical diseases. Its advantages include non-invasive, real-time dynamic imaging, low cost, etc., so it has been widely used in many fields such as cardiovascular, oncology, and obstetrics and gynecology. However, due to various factors (such as probe swing, patient position, performance of imaging equipment, etc.), ultrasonic images often have problems such as noise, artifacts, and signal loss, which affect their diagnostic accuracy. Therefore, how to improve the quality of ultrasonic images and enhance their diagnostic accuracy has become an important research topic in the field of medical imaging.

[0003] In recent years, deep learning technology has made remarkable progress in the field of image processing. Especially in medical image analysis, through its powerful feature learning ability, deep learning can automatically extract useful information in images, effectively improving the automation and intelligence level of image analysis. In ultrasonic image processing, deep learning can perform operations such as automatic recognition, noise reduction, restoration, feature extraction and classification on images, and can achieve comparable or even better results than traditional image processing methods.

[0004] However, there are still some technical bottlenecks and challenges in the current ultrasonic image analysis technology based on deep learning. First of all, due to the particularity of its imaging principle, ultrasonic 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 the image quality, so they cannot provide sufficiently accurate diagnostic basis. Secondly, existing ultrasonic image analysis methods often only consider a single image feature, 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 ultrasonic image analysis, combining information of multiple features, especially local texture, dynamic motion, and global structural features, has become the key to improving the performance of intelligent analysis of ultrasonic images. In addition, existing methods usually do not fully consider the individual differences of cases and fail to automatically adjust the weight distribution of features according to different case types, which makes some special types of cases unable to be effectively diagnosed.

[0005] In response to the above problems, more and more research has begun to propose comprehensive methods that use a variety of image processing techniques and deep learning models to perform multi-level and multi-dimensional processing on ultrasound images. For example, through a multi-level collaborative processing pipeline architecture, the images are preprocessed, including data cleaning, noise reduction, standardization and other processing steps, to improve the image quality from the source; at the same time, through deep learning models, image recognition and temporal analysis are performed on the images, multiple features such as texture, dynamics and structure are extracted, and these features are fused through an adaptive mechanism to further improve the accuracy and robustness of the diagnostic results.

[0006] In addition, with the continuous development of deep learning technology, the application of deep learning models such as convolutional neural networks, recurrent neural networks, and long short-term memory networks in ultrasound image analysis has gradually been recognized. These models can automatically learn features from a large amount of ultrasound image data, reduce manual intervention, and discover potential abnormal changes in complex images. Combining temporal analysis and spatial correlation analysis, deep learning models can capture the relationship between dynamic motion features and anatomical structure features, thus providing a more accurate basis for pathological diagnosis.

[0007] Generally speaking, the research on ultrasound image intelligent analysis technology is in rapid development, and the application of deep learning technology has brought new opportunities for the automated analysis of ultrasound images. 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 solved. Summary of the Invention

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

[0009] The above technical purpose of the present invention is achieved through the following technical solutions: An intelligent ultrasound image analysis system based on deep learning includes a data cleaning module, a dynamic noise reduction module, a spatial standardization module, an image recognition model, and a temporal analysis model. The specific operation process of the system includes the following steps: S100. Obtain the original ultrasound image data; among them, the original ultrasound image data includes an ultrasound scan sequence, and each ultrasound scan sequence contains multiple consecutive frames of grayscale images and corresponding spatio-temporal correlation information; S200. Preprocess the original ultrasound image data to generate standardized image data; among them, the preprocessing process is executed based on a multi-level collaborative processing pipeline architecture. The pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial standardization module, and real-time interaction and status synchronization are achieved between the modules through a data bus; S300. Input the standardized image data into the image recognition model, extract local texture features through the image recognition model, capture dynamic motion features using the time series analysis model, establish global structural features by combining spatial correlation analysis, fuse the above three types of features using an adaptive weight allocation mechanism, and dynamically adjust according to the case type. S400. Based on the fused local texture features, dynamic motion features, and global structural features, perform pathological classification and diagnostic decision-making.

[0010] By adopting the above technical solutions, based on the multi-level collaborative processing pipeline architecture of the data cleaning module, dynamic noise reduction module, and spatial normalization module, it is possible to effectively remove the noise and artifact areas in the image and repair the signal loss areas; this not only improves the clarity of the image, but also ensures the structural consistency of the image, enhancing the diagnostic value of the image; each module conducts real-time interaction and status synchronization through the data bus, ensuring the efficiency and seamless docking of the entire image processing process. The collaborative effect of steps such as data cleaning, noise reduction, and normalization enables the ultrasound image to reach a high quality standard in the preprocessing stage, laying a foundation for subsequent image recognition and diagnostic decision-making; by extracting local texture features through the image recognition model, capturing dynamic motion features with the time series analysis model, and establishing global structural features by combining spatial correlation analysis, the system can comprehensively consider various feature information, and the adaptive weight allocation mechanism dynamically adjusts the feature weights according to the case type, further improving the accuracy of pathological classification and diagnostic decision-making and enhancing the system's adaptability to different types of cases; then, by integrating local, dynamic, and global feature information, the system can comprehensively consider multi-dimensional features in the image during the diagnosis process, making the identification of the lesion area more accurate and capable of dealing with complex situations in the image, such as noise, artifacts, signal loss, etc., thus effectively improving the accuracy and robustness of the diagnosis; at the same time, the system can also automatically adjust the weight allocation of features according to the actual case type to ensure that the feature analysis of each case can be optimized according to its specific situation. This adaptive adjustment mechanism enables the system to make more accurate diagnoses when facing individual differences, avoiding misdiagnosis problems caused by insufficient standardization in traditional methods.

[0011] A further setting is that the step S200 is specifically: S201. Perform the following operations in the data cleaning module: Traverse all ultrasound scan sequences and detect whether there are artifact areas and signal loss areas in the consecutive frames of each ultrasound scan sequence. If it is detected that a certain frame in the ultrasound scan sequence has an artifact area with an area ratio exceeding the set threshold, activate the adjacent frame compensation mechanism and use the median gray value of the same position area of the previous and next 5 frames of images in this ultrasound scan sequence for filling. If signal loss regions appear in 3 or more consecutive frames of the ultrasound scan sequence, based on the probe movement trajectory data in the spatio-temporal correlation information, an elastic interpolation algorithm is used to reconstruct the pixel matrix of the missing region; S202. Perform multi-scale feature fusion denoising in the dynamic denoising module, specifically including: Perform wavelet packet decomposition on the grayscale image of each ultrasound scan sequence to generate a high-frequency subband coefficient matrix and a low-frequency subband coefficient matrix; Adopt a dual-threshold denoising strategy based on adaptive adjustment of the signal-to-noise ratio for processing: For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, apply the Bayesian shrinkage function for noise suppression; when the local signal-to-noise ratio is higher than the set second threshold, keep the original coefficients without processing; Perform non-local means filtering on the low-frequency subband coefficient matrix; S203. Implement multi-modal registration in the spatial normalization module, specifically including: Construct an anatomical structure recognition model based on a deep neural network, input the preprocessed ultrasound image data into the anatomical structure recognition model, extract the anatomical features of the organ edges and blood vessel directions through the convolutional layer, and classify and determine whether it is a standard section through the fully connected layer; If it is a standard section, directly perform subsequent processing; if it is a non-standard section, trigger the elastic registration process and then perform subsequent processing; Perform the following operations on the registered ultrasound image data: Unify the spatial resolution through bicubic interpolation; Linearly map the original grayscale value to the 8-bit range and perform clipping processing on the overexposed area; S204. Integrate the processing results of the data cleaning module, the dynamic denoising module, and the spatial normalization module, and output the standardized image data that meets the spatial resolution, grayscale range, and anatomical consistency.

[0012] By adopting the above technical solutions, the data cleaning module can identify the missing areas and artifacts in the images, and supplement them through the interpolation and restoration techniques of similar image areas to ensure the integrity of the images; the dynamic noise reduction module adopts a noise reduction algorithm based on convolutional neural network, which can effectively remove the noise in the images while retaining the detailed information and avoiding the loss of valuable structural information during the denoising process; the spatial normalization module makes the image data collected at different times and by different devices have a consistent spatial standard through an automated registration method, thus realizing the unification of multi-source image data; the above process greatly improves the image quality and maintains the integrity of the important detailed information in the images. The optimization of the denoising effect makes the subtle lesions more obvious. Especially in ultrasonic images, due to the influence of noise, the lesion areas are often obscured. The addition of this module effectively improves the clarity and identifiability of the images.

[0013] A further setting is that in the step S201, the specific method for reconstructing the pixel matrix body of the missing area by using the elastic interpolation algorithm is as follows: When it is detected that there are signal loss areas in 3 or more consecutive frames of the ultrasonic scanning sequence, lock the loss time window and extract the discrete coordinate data set corresponding to the corresponding time period; Fit the discrete coordinate data set into a continuous motion trajectory curve through B-spline curve fitting, eliminate the coordinate mutation caused by operation jitter, and generate a high-precision probe motion path; Select anatomical feature points as control points in the normal frames before and after the signal loss, and establish a spatio-temporal mapping table of time-probe coordinates-anatomical point coordinates; Based on the spatio-temporal mapping table, model the ultrasonic scanning area as an elastic thin plate model, set the regional stiffness parameter, and bind the anatomical control points as thin plate anchor points; Perform coarse-grained structure restoration and fine-grained detail optimization, and combine the elastic thin plate model constraint with the verification of the medical rule library to generate the restored image sequence.

[0014] By adopting the above technical solutions, traditional image restoration methods often rely on simple linear interpolation or nearest neighbor interpolation methods. Although these methods are simple, they often cause obvious artifacts or distortion in the restored areas for complex ultrasonic images; while the elastic interpolation algorithm can accurately reconstruct the missing areas through adaptive curve fitting technology, and naturally fuse with the structure and texture of the surrounding areas, avoiding information loss or over-smoothing in image restoration, and the restored images have high visual continuity and consistency; especially when dealing with the missing areas caused by equipment vibration or patient body position change, the elastic interpolation algorithm shows high restoration quality and can effectively restore the important structural information in the images, improving the integrity and credibility of the images.

[0015] A further setting is that in the step S203, the specific process of triggering the elastic registration process is as follows: Based on the preprocessed ultrasonic image data and the associated spatio-temporal information, construct a multi-resolution B-spline deformation field; Match the overall shape at low resolution, adjust local details at high resolution, and adjust the regularization parameter.

[0016] By adopting the above technical solution, the spatio-temporal registration technology can accurately align each frame of images 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 spatio-temporal information, and use the elastic deformation field for precise registration; this registration method not only considers the spatial dimension of the images, but also makes full use of the information in the time dimension to ensure the global consistency of the images; at the same time, during the registration process, the alignment between images is gradually optimized through the multi-resolution method, ensuring the precise registration effect at high resolution, thereby improving the detail retention of the images and avoiding structural distortion caused by registration errors.

[0017] A further setting is that in the step S300, the specific process of extracting local texture features through the image recognition model, capturing dynamic motion features using the time series analysis model, and establishing global structural features by combining spatial correlation analysis is as follows: Use the pre-trained image recognition model to analyze the standardized image data and automatically identify the local detail features therein, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity; Through the time series analysis model, perform dynamic tracking on the continuous frame sequence in the standardized image data to quantify the dynamic motion features; Divide the single-frame images in the standardized image data into grids, and model the organ position relationship and the vascular network topology structure through spatial correlation analysis to generate global structural features representing the anatomical structure.

[0018] By adopting the above technical solution, multi-level feature information can be extracted from the images. The image recognition model focuses on capturing local texture features and can accurately identify subtle lesion signs, such as the edges of masses, changes in blood flow, etc.; the time series analysis model captures the dynamic features during the movement process through time series data analysis, such as the movement of organs, changes in tumors, etc.; spatial correlation analysis combines the global structural information to identify possible lesion areas from a large range; the comprehensive extraction of these features enables the system to comprehensively analyze ultrasonic images and identify more detailed lesion areas. The combination of local features and global structures makes the system more accurate in the diagnosis of lesions and avoids missed diagnoses caused by single-feature analysis.

[0019] A further setting is that in the step S300, an adaptive weight allocation mechanism is adopted to fuse the above three types of features, and the feature weight ratio is dynamically adjusted according to the case type. Specifically: Perform standardization processing on the local texture features, dynamic motion features, and global structure features respectively to eliminate the differences in numerical dimensions; Configure the initial weights of the local texture features, dynamic motion features, and global structure features; Dynamically adjust the weight ratio according to the actual case type: Adopt temperature scaling technology to perform probability calibration on the fused features and optimize the classification confidence.

[0020] By adopting the above technical solution, the local texture features, dynamic motion features, and global structure features each carry different information in the image data. For example, the local texture features describe the microscopic structure in the image, the dynamic motion features reflect the dynamic changes of organs and lesion areas, and the global structure features provide a macroscopic control of the entire anatomical structure; the introduction of the 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 the accuracy of diagnosis but also reduces the over-reliance on specific features and avoids misdiagnosis or missed diagnosis caused by a single feature. Through the standardization processing of the three types of features, the possible differences in numerical dimensions between them are eliminated, ensuring that they have relatively balanced influence when fusing features. After feature fusion, the fused features are subjected to probability calibration by adopting temperature scaling technology, thereby further optimizing the classification confidence.

[0021] A further setting is that the step S400 is specifically: S401. Receive the feature data fused in step S300 and simultaneously perform multi-task diagnosis: Automatically detect the 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 the dynamic motion features and generate a motion trajectory map that changes over time; Establish an organ position relationship model through spatial grid division to identify abnormal blood vessel directions or tissue deformations; S402. Perform hierarchical verification and correction; S403. Generate a visual diagnostic report; S404. Establish a self-optimizing system; S405. Output the final structured report.

[0022] By adopting the above technical solution, in step S401, the system not only automatically detects suspicious lesion areas in the standardized image data, but also verifies the rationality of the anatomical position in combination with the standard section data. This ability enables the system to judge whether there are lesion areas based on more accurate anatomical structure information and provide instant feedback in the actual clinical environment, helping doctors quickly confirm the lesion site and make 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 orientation, etc.

[0023] A further setting is that the specific step S402 is as follows: Compare the preliminary result with the preset medical standard; if a contradiction is found, automatically trigger the multi-modal registration process in step S203 to recalibrate the data; for cases with doubtful diagnoses, return to step S300 to adjust the weight assignment 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 in step S201 to repair the data for the mutation result.

[0024] By adopting the above technical solution, in step S402, when there is a contradiction between the preliminary diagnosis result and the preset medical standard, the system can automatically trigger the multi-modal registration process for calibration; this process effectively avoids diagnostic errors caused by image deviation or error, ensures the accuracy of the ultrasound image data. By checking 5 consecutive frames of image data, the system can confirm the consistency of the diagnostic conclusions, thus avoiding the risk of misdiagnosis of a single frame; for doubtful cases, the system will return to step S300 according to the scheduling mechanism to adjust the weight assignment mechanism and conduct a secondary analysis again. This link provides a second chance for diagnosis, ensuring that the system can provide more accurate analysis results when facing complex lesions. Through this mechanism, the error rate of the system is significantly reduced, further enhancing its reliability in clinical practice; through the comparative analysis of 5 consecutive frames of ultrasound image data, the system can confirm the consistency of the diagnostic results and compensate and repair the mutation results. This mechanism enhances the adaptability of the model when facing dynamic lesions or instantaneous changes, ensuring the real-time tracking and accurate diagnosis of the patient's condition.

[0025] A further setting is that the specific step S403 is as follows: Convert the dynamic motion characteristics extracted in step S300 into a motion curve graph, and generate a clear frequency spectrum graph in combination with the data after noise reduction in step S202. Overlay semi-transparent color blocks on the standardized image data to mark the lesion area, and adopt the HSV color gamut mapping technology to convert the diagnostic confidence into hue and saturation parameters. Automatically match the diagnostic template according to the analysis results and highlight the key judgment basis.

[0026] By adopting the above technical solution, by converting the dynamic motion features into a motion curve graph and combining the data after noise reduction to generate a clear spectrogram, the system can intuitively present the change trend and dynamic features of the lesion area. Doctors can quickly judge the type, progression of the lesion and its relationship with other organs by observing these graphs, thus accelerating the diagnostic decision-making; overlaying semi-transparent color blocks on the standardized image data to mark the lesion area, and using the HSV color gamut mapping technology to convert the diagnostic confidence into hue and saturation parameters. This color block marking method not only makes the lesion area more prominent, but also can intuitively reflect the confidence level of the system in the diagnostic result, helping doctors to be more clear and rapid in decision-making.

[0027] A further setting is that the step S404 is specifically: Collect the difference data between the clinical diagnosis result 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 backpropagation algorithm; When the misjudgment rate of a certain type of case exceeds the set threshold, automatically optimize the initial weight ratio of the local texture features, dynamic motion features and global structure features in step S300; Regularly transmit the optimized parameters back to steps S201 to S203 to improve the artifact repair and image multi-modal registration accuracy.

[0028] By adopting the above technical solution, the system can collect the difference data between the clinical diagnosis result and the multi-task diagnosis through the backpropagation algorithm, and synchronously update the parameters of the image recognition model, the time series analysis model and the adaptive weight allocation mechanism. This process enables the system to automatically adjust and optimize its algorithm according to the new clinical data and diagnostic situation, ensuring high diagnostic accuracy when facing new types of 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 over time, the diagnostic ability of the system is continuously strengthened; the subsequent feedback mechanism ensures that the ultrasound image can still provide high-quality data input in various complex situations, avoiding misdiagnosis and missed diagnosis caused by poor image quality.

[0029] In summary, the present invention has the following beneficial effects: The intelligent analysis system for ultrasound images based on deep learning significantly improves the quality of images and the accuracy of diagnosis through multi-level and multi-dimensional image processing and analysis technologies, and has high clinical application value. Brief Description of the Drawings

[0030] Figure 1 Schematic diagram of the process of the embodiment. Specific implementation manners

[0031] The present invention will be further described in detail below with reference to the accompanying drawings.

[0032] As shown in the Figure 1 accompanying drawings; This embodiment discloses an intelligent ultrasonic image analysis system based on deep learning, including a data cleaning module, a dynamic noise reduction module, a spatial normalization module, an image recognition model, and a temporal analysis model. The specific process of the system operation includes the following steps: S100. Obtain the original ultrasonic image data; wherein, the original ultrasonic image data includes ultrasonic scan sequences, and each ultrasonic scan sequence contains multiple consecutive frames of grayscale images and corresponding spatio-temporal correlation information; S200. Preprocess the original ultrasonic image data to generate normalized image data; wherein, the preprocessing process is executed based on a multi-level collaborative processing pipeline architecture, and the pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial normalization module. Real-time interaction and status synchronization are achieved between modules through a data bus; S300. Input the normalized image data into the image recognition model, extract local texture features through the image recognition model, capture dynamic motion features using the temporal analysis model, establish global structural features by combining spatial correlation analysis, fuse the above three types of features using an adaptive weight allocation mechanism, and dynamically adjust according to the case type; S400. Perform pathological classification and diagnostic decision based on the fused local texture features, dynamic motion features, and global structural features.

[0033] Among them, step S200 is specifically as follows: S201. Perform the following operations in the data cleaning module: Traverse all ultrasonic scan sequences, and detect whether there are artifact regions and signal loss regions in the consecutive frames of each ultrasonic scan sequence; If it is detected that a certain frame in the ultrasonic scan sequence has an artifact region with an area ratio exceeding a set threshold, and the set threshold is 15%; then activate the adjacent frame compensation mechanism, and use the median value of the gray-scale values of the same-position regions of the previous and next 5 frames of images in the ultrasonic scan sequence for filling; If it is detected that the signal loss region appears in 3 or more consecutive frames of the ultrasonic scan sequence, based on the probe movement trajectory data in the spatio-temporal correlation information, an elastic interpolation algorithm is used to reconstruct the missing region pixel matrix; S202. Perform multi-scale feature fusion noise reduction in the dynamic noise reduction module, specifically including: Perform wavelet packet decomposition on the grayscale images of each ultrasonic scanning sequence to generate a high-frequency subband coefficient matrix and a low-frequency subband coefficient matrix; Adopt a dual-threshold denoising strategy based on adaptive adjustment of signal-to-noise ratio for processing: For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, apply the Bayesian shrinkage function for noise suppression; when the local signal-to-noise ratio is higher than the set second threshold, retain the original coefficients without processing; the set first threshold is 8 dB, and the set second threshold is 15 dB; Perform non-local means filtering on the low-frequency subband coefficient matrix, specifically by setting the search window radius to 7 pixels, the similarity window radius to 3 pixels, and the Gaussian weighting parameter to 0.6; S203. Implement multimodal registration in the spatial normalization module, specifically including: Construct an anatomical structure recognition model based on a deep neural network, input the preprocessed ultrasonic image data into the anatomical structure recognition model, extract the anatomical features of organ edges and blood vessel directions through the convolutional layer, and classify and determine whether it is a standard section through the fully connected layer; If it is a standard section, directly perform subsequent processing; if it is a non-standard section, trigger the elastic registration process and then perform subsequent processing; Perform the following operations on the registered ultrasonic image data: Unify the spatial resolution to 0.2 mm × 0.2 mm through bicubic interpolation; Linearly map the original grayscale value to the 8-bit range and perform clipping processing on the overexposed area; S204. Integrate the processing results of the data cleaning module, the dynamic denoising module, and the spatial normalization module, and output the standardized image data that meets the spatial resolution, grayscale range, and anatomical consistency.

[0034] Among them, in step S201, the specific method of reconstructing the missing area pixel matrix body using the elastic interpolation algorithm is: When it is detected that there are signal loss areas in 3 or more consecutive frames of the ultrasonic scanning sequence, lock the loss time window and extract the discrete coordinate data set corresponding to the corresponding time period; the determination condition for the signal loss area is: the standard deviation of the grayscale values in the same anatomical area in 3 consecutive frames of the ultrasonic scanning sequence is lower than 5% of the normal value.

[0035] Fit the discrete coordinate data set into a continuous motion trajectory curve through B-spline curve fitting, eliminate the coordinate mutation caused by operation jitter, and generate a high-precision probe motion path; Among them, the B-spline curve fitting is specifically: Perform cubic B-spline interpolation on the discrete coordinates to generate a continuous trajectory curve; When the distance mutation between adjacent coordinate points exceeds 2 mm, start the trajectory smoothing correction algorithm to remove abnormal points and refit.

[0036] Select anatomical feature points as control points in the normal frames before and after signal loss, and establish a spatio-temporal mapping table of time-probe coordinates-anatomical point coordinates; Among them, the selection of anatomical feature points includes: Static control points: bone marker points, blood vessel bifurcation center points, whose coordinate change rate ≤ 0.1 mm / frame; Dynamic control points: the thinnest points of the myocardial wall, the center points of valve opening and closing, and the angle between their movement directions and the tangent direction of the probe trajectory ≤ 15°.

[0037] Based on the spatio-temporal mapping table, model the ultrasonic scanning area as an elastic thin plate model, set the regional stiffness parameter, and bind the anatomical control points as thin plate anchor points; The parameters of the elastic thin plate model are set such that the stiffness coefficient of the myocardial region is 10 times that of the blood flow region. The deformation constraint conditions of the thin plate include: the myocardial stretching limit ≤ 0.2 mm / frame, and the blood vessel branch angle change ≤ 5° / frame.

[0038] Perform coarse-grained structure restoration and fine-grained detail optimization, combine the elastic thin plate model constraints with the verification of the medical rule base to generate a repaired image sequence. Among them, the coarse-grained structure restoration process is to project the anatomical structures of adjacent 5 frames along the trajectory to the lost frame based on the elastic thin plate model to generate an initial gray distribution; the fine-grained detail optimization process is to perform gradient sharpening and dynamic blur addition on the tissue boundary under the thin plate stiffness constraint.

[0039] Among them, in step S203, the specific process of triggering the elastic registration process is: Based on the preprocessed ultrasonic image data and associated spatio-temporal information, construct a multi-resolution B-spline deformation field; Match the overall shape at low resolution, the grid spacing of the deformation field 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, adjust the local details and adjust the regularization parameter. The grid spacing of the deformation field is reduced to 4 mm, and the value of the regularization parameter will decrease. The purpose is to correct details such as the myocardial wall thickness and blood vessel bifurcation angle.

[0040] Among them, in step S300, the specific process of extracting local texture features through the image recognition model, capturing dynamic motion features using the time series analysis model, and establishing global structure features by combining spatial correlation analysis is: Use a pre-trained image recognition model, adopt the ResNet-50 network, analyze the standardized image data, and automatically identify the local detail features therein, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity; specifically: Adopt a hierarchical progressive feature extraction strategy, capture basic texture patterns through low-level convolutions of the ResNet-50 network, and identify complex pathological patterns through high-level convolutions; Implement multi-scale feature fusion, upsample feature maps at different levels to a unified resolution and then perform channel splicing; finally, by embedding a channel attention mechanism, automatically suppress noise channels and enhance the response weights of lesion-related channels.

[0041] Through a time series analysis model, using an LSTM network, dynamically track the sequence of consecutive frames in the standardized image data, and quantify dynamic motion features such as myocardial motion rate and valve opening and closing cycles; Divide a single-frame image in the standardized image data into a grid, divided into 16×16 grid blocks, and model the organ position relationship and vascular network topology through spatial correlation analysis to generate global structural features representing anatomical structures; specifically: Perform 16×16 grid division on the standardized single-frame image to generate 256 local blocks; Attach spatial position encoding to each block to record its coordinate offset relative to the anatomical landmark point; Calculate the correlation weights between blocks through a multi-head self-attention mechanism, and focus on strengthening the following medically significant regions: The matching relationship in the ventricular symmetry region; The connectivity path of the vascular bifurcation point; The continuity features of the tissue boundary.

[0042] Among them, in step S300, an adaptive weight allocation mechanism is used to fuse the above three types of features, and the feature weight ratio is dynamically adjusted according to the case type. Specifically: Perform standardization processing on the local texture features, dynamic motion features, and global structural features respectively to eliminate the difference in numerical dimensions; Configure the initial weights of the local texture features, dynamic motion features, and global structural features; The initial weights are set as: The weight of local texture features is 40%; The weight of dynamic motion features is 30%; The weight of global structural features is 30%; Dynamically adjust the weight ratio according to the actual case type: For example, when dealing with static lesions such as calcified plaques, the weight of local texture features is increased to 60%; When analyzing dynamic abnormalities such as heart valve regurgitation, the weight of dynamic motion features is increased to 50%; Use temperature scaling technology to perform probability calibration on the fused features and optimize the classification confidence.

[0043] Among them, step S400 is specifically as follows: S401. Receive the fused feature data in step S300, and simultaneously perform multitask diagnosis: 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 the dynamic motion characteristics and generate a motion trajectory diagram that changes over time; Establish an organ position relationship model through spatial grid division to identify abnormal blood vessel directions or tissue deformations; S402. Perform hierarchical verification and correction; S403. Generate a visual diagnostic report; S404. Establish a self-optimizing system; S405. Output the final structured report.

[0044] Among them, step S402 is specifically as follows: Compare the preliminary results with the preset medical standards, such as the normal range of ventricular systolic velocity; if contradictions are found, automatically trigger the multimodal registration process in step S203 to recalibrate the data; for cases with doubtful diagnoses, return to step S300 to adjust the weight distribution mechanism, such as increasing the local texture weight from 40% to 60% for secondary analysis; Check whether the diagnostic conclusions of 5 consecutive frames of ultrasound image data are consistent, and call the adjacent frame compensation mechanism in step S201 to repair the data for mutation results.

[0045] Among them, step S403 is specifically as follows: Convert the dynamic motion characteristics extracted in step S300 into a motion curve diagram, and generate a clear spectrogram in combination with the data after noise reduction in step S202; Overlay semi-transparent color blocks on the standardized image data to mark the lesion areas, and use the HSV color gamut mapping technology to convert the diagnostic confidence into hue and saturation parameters, where the red series indicates a confidence level > 85%, the yellow series indicates 50% - 85%, and the blue series indicates < 50%; Automatically match the diagnostic template according to the analysis results and highlight the key judgment bases.

[0046] Among them, step S404 is specifically as follows: Collect the difference data between the clinical diagnosis results and the multitask 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 ratio coefficients in the adaptive weight distribution mechanism; When the misjudgment 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 of step S202 is adjusted in conjunction; The optimized parameters are regularly fed back to steps S201 to S203 to improve the accuracy of artifact restoration and multi-modal image registration.

[0047] Application Example 1 In step S100, ultrasound image data of 200 patients with suspected mitral regurgitation were obtained from the case database of a tertiary hospital, each case containing a 30-second continuous ultrasound scan sequence, with a total data volume of 180,000 frames; 12% of the frames in the original ultrasound image data were detected to have artifacts, with an average area of 18%, 8% of the ultrasound scan sequences had 3-5 consecutive frame signal loss, and the grayscale standard deviation of the myocardial area decreased from the normal value of 45±8 to 2±1; In step S201, the adjacent frame compensation mechanism is activated, and the image structure similarity index is increased from 0.72 to 0.89 after the artifact area is filled with the median of the adjacent 5 frames; When the elastic interpolation algorithm reconstructs the signal loss area, the cubic B-spline curve is used to fit the probe trajectory, combined with the elastic thin plate model. The elastic thin plate model is configured with a myocardial stiffness coefficient of 120kPa and a blood flow area of 12kPa. The error between the reconstructed valve motion trajectory and the gold standard MRI is 0.8mm, which is less than the clinically allowable error of 1.5mm.

[0048] In step S202, the db4 basis function is used for wavelet packet decomposition, and the high-frequency subband Bayesian shrinkage is used to increase the contrast-to-noise ratio of the myocardial boundary from 12 dB to 24 dB; Low-frequency non-local mean filtering improves tissue texture clarity by 37% and optimizes computational time to 18ms / frame.

[0049] In step S203, the anatomical structure recognition model accurately identifies the standard four-chamber heart section, and implements multi-resolution registration for non-standard sections, 16mm grid coarse registration + 4mm grid fine registration. After registration, the ventricular volume measurement error is reduced from 9.2ml to 2.1ml.

[0050] 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.

[0051] 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; The spatial meshing modeled the symmetry of the ventricles, and the sensitivity of detecting abnormal areas of left atrial / left ventricular volume ratio in regurgitation cases reached 92%; 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, significantly better than that of the single-feature model, which ranges from 0.82 to 0.89.

[0052] In step S400, 5 false negatives were found through hierarchical verification, triggering weight adjustment. That is, after the local texture weight was increased from 40% to 55%, the detection rate was increased to 98.5%. The self-optimizing system updated the model parameters 12,000 times in total, reducing the prediction error of myocardial motion trajectories by 42%. The final diagnostic accuracy rate is 91.3%, and the diagnostic time is shortened from the traditional 45 minutes to 8 minutes.

[0053] Application Example 2 In step S100, ultrasonic data of 350 patients with chronic liver diseases were obtained from the case database of a certain tertiary hospital. Each case included the dynamic sequence of the long axis of the portal vein. 18% of the original ultrasonic images had rib artifacts, with an area ratio of 23%, and the continuous frame signal loss rate at the portal vein branch was 9%.

[0054] In step S201, when reconstructing the portal vein branch using the elastic interpolation algorithm, the vascular stiffness coefficient was set to 80 kPa, and the reconstructed vascular diameter error was 0.3 mm, which was less than the clinical standard of 0.5 mm. In step S202, the contrast-to-noise ratio of the hepatic lobule structure was increased from 18 dB to 31 dB by high-frequency sub-band double-threshold noise reduction; low-frequency non-local mean filtering retained the hepatic portal sheath structure, and the texture uniformity index was increased by 29%. In step S203, after registration, the standard deviation of the angle of the right hepatic vein decreased from 7.2° to 1.8°, and the spatial resolution was unified to 0.2 mm / px, enabling the measurement accuracy of the fibrous septum to reach 0.05 mm. In step S300, the image recognition model identified the reticular changes in the liver parenchyma, and the specificity for fibrosis above F3 was 91%. The time series analysis model captured the attenuation of the portal vein pulsation. The pulsatility index changed from 0.38 to 0.22, and the predicted error of the hepatic vein pressure gradient was 3.2 mmHg. Adaptive weight allocation focuses on local texture features, increasing to 60%. The accuracy rate of fibrosis staging of the fusion model and the overall Kappa value is 0.89, significantly better than that of ultrasonic elastography. In step S400, 32 cases of portal vein thrombosis were detected by multi-task detection, with a sensitivity of 95%, and the range of thrombosis was accurately located through HSV color gamut mapping. Compared with the gold standard of liver biopsy, the diagnostic coincidence rate is 89.4%, and the diagnostic time is shortened from 3 days to real-time analysis.

[0055] The above two embodiments verify the excellent performance of the method in dynamic heart diagnosis and static liver lesion analysis. All indicators exceed the clinical requirements, confirming the effectiveness.

[0056] This specific embodiment is only an interpretation of the present invention and is not a limitation thereof. After reading this specification, those skilled in the art can make modifications to this embodiment that do not contribute creatively as needed, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. An intelligent ultrasonic image analysis system based on deep learning, characterized in that 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 specific process of the system operation includes the following steps: S100. Obtain the original ultrasound image data. Among them, the original ultrasound image data includes an ultrasound scan sequence, and each ultrasound scan sequence contains multiple consecutive frames of grayscale images and corresponding spatio-temporal correlation information; S200. Preprocess the original ultrasound image data to generate normalized image data. Among them, the preprocessing process is executed based on a multi-level collaborative processing pipeline architecture. The pipeline architecture includes a data cleaning module, a dynamic noise reduction module, and a spatial normalization module, and real-time interaction and status synchronization are achieved between modules through a data bus; S300. Input the normalized 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, establish global structural features by combining spatial correlation analysis, fuse the above three types of features using an adaptive weight allocation mechanism, and dynamically adjust according to the case type; S400. Based on the fused local texture features, dynamic motion features, and global structural features, perform pathological classification and diagnostic decision-making.

2. The ultrasound image intelligent analysis system based on deep learning according to claim 1, wherein: The specific step S200 is as follows: S201. Perform the following operations in the data cleaning module: Traverse all ultrasound scan sequences, and detect whether there are artifact regions and signal loss regions in consecutive frames of each ultrasound scan sequence; If it is detected that a certain frame in the ultrasound scan sequence has an artifact region with an area ratio exceeding the set threshold, activate the adjacent frame compensation mechanism, and use the median gray value of the same position region of the previous and next 5 frames of images in the ultrasound scan sequence for filling; If it is detected that the signal loss region appears in 3 or more consecutive frames of the ultrasound scan sequence, based on the probe motion trajectory data in the spatio-temporal correlation information, use the elastic interpolation algorithm to reconstruct the missing region pixel matrix; S202. Perform multi-scale feature fusion noise reduction in the dynamic noise reduction module, specifically including: Perform wavelet packet decomposition on the grayscale image of each ultrasound scan sequence to generate a high-frequency subband coefficient matrix and a low-frequency subband coefficient matrix; Adopt a dual-threshold noise reduction strategy based on adaptive adjustment of the signal-to-noise ratio for processing: For the high-frequency subband coefficient matrix, when the local signal-to-noise ratio is lower than the set first threshold, apply the Bayesian shrinkage function for noise suppression; when the local signal-to-noise ratio is higher than the set second threshold, keep the original coefficient without processing; Perform non-local mean filtering on the low-frequency subband coefficient matrix; S203. Implement multi-modal registration in the spatial normalization module, specifically including: Construct an anatomical structure recognition model based on a deep neural network, input the preprocessed ultrasound image data into the anatomical structure recognition model, extract anatomical features of organ edges and blood vessel directions through the convolutional layer, and classify and determine whether it is a standard section through the fully connected layer; If it is a standard section, directly perform subsequent processing; if it is a non-standard section, trigger the elastic registration process and then perform subsequent processing; Perform the following operations on the registered ultrasound image data: Unify the spatial resolution through bicubic interpolation; Linearly map the original gray value to the 8-bit range, and perform clipping processing on the overexposed region; S204. Integrate the processing results of the data cleaning module, the dynamic noise reduction module, and the spatial normalization module, and output the standardized image data that meets the spatial resolution, gray scale range, and anatomical consistency.

3. The intelligent ultrasonic image analysis system based on deep learning according to claim 2, wherein: In the step S201, the specific method for reconstructing the missing region pixel matrix body using the elastic interpolation algorithm is as follows: When it is detected that there are signal loss regions in 3 or more consecutive frames of the ultrasonic scanning sequence, lock the loss time window and extract the discrete coordinate data set corresponding to the corresponding time period; Fit the discrete coordinate data set into a continuous motion trajectory curve through B-spline curves, eliminate the coordinate mutations caused by operation jitter, and generate a high-precision probe motion path; Select anatomical feature points as control points in the normal frames before and after the signal loss, and establish a spatio-temporal mapping table of time - probe coordinates - anatomical point coordinates; Based on the spatio-temporal mapping table, model the ultrasonic scanning area as an elastic thin plate model, set the regional stiffness parameter, and bind the anatomical control points as thin plate anchor points; Perform coarse-grained structure restoration and fine-grained detail optimization, and combine the elastic thin plate model constraint and the medical rule base verification to generate the repaired image sequence.

4. The intelligent ultrasonic image analysis system based on deep learning according to claim 3, wherein: In the step S203, the specific method for triggering the elastic registration process is as follows: Based on the preprocessed ultrasonic image data and the associated spatio-temporal information, construct a multi-resolution B-spline deformation field; Match the overall shape at low resolution and adjust the local details at high resolution, and adjust the regularization parameter.

5. The intelligent ultrasonic image analysis system based on deep learning according to claim 4, characterized in that: In the step S300, the specific method for extracting local texture features through the image recognition model, capturing dynamic motion features using the time series analysis model, and establishing global structure features by combining spatial correlation analysis is as follows: Use the pre-trained image recognition model to analyze the standardized image data, and automatically identify the local detail features therein, including tissue microstructure, abnormal plaque distribution, and tissue boundary clarity; Through the time series analysis model, perform dynamic tracking on the continuous frame sequence in the standardized image data to quantify the dynamic motion features; Divide the single-frame image in the standardized image data into grids, and model the organ position relationship and the vascular network topology structure through spatial correlation analysis to generate global structure features representing the anatomical structure.

6. The intelligent ultrasonic image analysis system based on deep learning according to claim 5, wherein: In the step S300, adopt an adaptive weight allocation mechanism to fuse the above three types of features, and dynamically adjust the feature weight ratio according to the case type. The specific method is as follows: Perform standardized processing on the local texture features, dynamic motion features, and global structure features respectively to eliminate the numerical dimension differences; Configure the initial weights of the local texture features, dynamic motion features, and global structure features; Dynamically adjust the weight ratio according to the actual case type: Adopt the temperature scaling technology to perform probability calibration on the fused features and optimize the classification confidence.

7. The intelligent ultrasonic image analysis system based on deep learning according to claim 6, characterized in that: The step S400 is specifically as follows: S401. Receive the feature data fused in the step S300, and simultaneously perform multi-task diagnosis: Automatically detect the 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 the step S203; Analyze the dynamic motion features and generate a motion trajectory map that changes with time; Establish an organ position relationship model through spatial grid division, and identify abnormal blood vessel directions or tissue deformations; S402. Hierarchical verification and correction; S403. Generate a visual diagnostic report; S404. Establish a self-optimizing system; S405. Output the final structured report.

8. The intelligent ultrasonic image analysis system based on deep learning according to claim 7, characterized in that: The specific steps of step S402 are as follows: Compare the preliminary results with the preset medical standards; if contradictions are found, automatically trigger the multi-modal registration process in step S203 to recalibrate the data; for cases with doubtful diagnoses, return to step S300 to adjust the weight assignment mechanism for secondary analysis; Check whether the diagnostic conclusions of 5 consecutive frames of ultrasonic image data are consistent, and call the adjacent frame compensation mechanism in step S201 to repair the data for mutant results.

9. The intelligent ultrasonic image analysis system based on deep learning according to claim 8, wherein: The specific steps of step S403 are as follows: Convert the dynamic motion features extracted in step S300 into a motion curve graph, and generate a clear spectrogram by combining the data after noise reduction in step S202; Overlay semi-transparent color blocks on the standardized image data to mark the lesion area, and use the HSV color gamut mapping technology to convert the diagnostic confidence into hue and saturation parameters; Automatically match the diagnostic template according to the analysis results and highlight the key judgment bases.

10. The intelligent ultrasonic image analysis system based on deep learning according to claim 9, characterized in that: The specific steps of step S404 are 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 assignment mechanism through the backpropagation algorithm; When the misjudgment rate of a certain type of case exceeds the set threshold, automatically optimize the initial weight ratio of the local texture features, dynamic motion features, and global structure features in step S300; Regularly transmit the optimized parameters back to steps S201 to S203 to improve the artifact repair and image multi-modal registration accuracy.

Images