Ultrasonic image data classification method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based ultrasonic image data classification method and system, and the method comprises the steps: firstly obtaining a real-time ultrasonic scanning signal sequence containing the time sequence change characteristics of a tissue elastic parameter and a hemodynamic parameter, carrying out the noise suppression and motion artifact compensation processing, generating a standardized ultrasonic image sequence, and marking the coordinates of an anatomical boundary; then performing multi-scale anatomical structure decomposition on the ultrasonic image to obtain a local feature map set of different organization levels, inputting the local feature map set into a cascade deep classification network, and realizing cross-frame feature fusion and dynamic weight adjustment through spatial-temporal feature alignment and a multi-granularity attention distribution module to obtain a spatial-temporal feature fusion model; and the abnormal region classification probability distribution and the spatial topological relation graph are output, finally, a multi-modal diagnosis report is generated according to the abnormal region classification probability distribution and the spatial topological relation graph, an interactive three-dimensional visual interface containing risk level labels and treatment suggestions is generated after the multi-modal diagnosis report is compared with historical cases, and ultrasonic image classification accuracy and diagnosis efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to an artificial intelligence-based ultrasonic image data classification method and system. Background Art

[0002] In the field of medical ultrasonic imaging, accurately classifying ultrasonic image data is crucial for disease diagnosis and treatment.

[0003] Most of the existing ultrasonic image classification methods only analyze static ultrasonic images, ignoring the temporal variation characteristics of elastic parameters and hemodynamic parameters of tissues during the dynamic process, resulting in the inability to comprehensively capture the physiological and pathological information of tissues, and limiting the accuracy and reliability of diagnostic results.

[0004] Regarding the problems of noise and motion artifacts commonly existing in ultrasonic images, the processing effect of traditional methods is limited, and it fails to effectively eliminate their impact on image quality, thereby interfering with subsequent image analysis and feature extraction. In terms of feature extraction, existing technologies often only focus on single-level image features, making it difficult to deeply explore the complex relationships between different tissue levels and unable to comprehensively display the rich information contained in ultrasonic images. In the classification model, traditional classification networks lack effective cross-frame feature fusion mechanisms and dynamic weight adjustment capabilities, making it difficult to process the complex spatio-temporal information in dynamic ultrasonic images, resulting in inaccurate classification of abnormal regions. Moreover, the existing diagnostic report generation methods are relatively single, usually only simply describing image features, lacking the integration of multi-modal information, and rarely making effective comparisons with historical cases, unable to provide intuitive and comprehensive diagnostic assistance for doctors, and difficult to meet the requirements of efficient and accurate diagnosis in clinical practice. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide an artificial intelligence-based ultrasonic image data classification method, and the method includes:

[0006] Obtain a real-time ultrasonic scanning signal sequence of a target object, wherein the real-time ultrasonic scanning signal sequence includes multiple frames of dynamic ultrasonic images, and the temporal variation characteristics of tissue elastic parameters and hemodynamic parameters are embedded in each frame of dynamic ultrasonic image;

[0007] Perform noise suppression and motion artifact compensation processing on the real-time ultrasonic scanning signal sequence to generate a denoised and normalized ultrasonic image sequence, and mark the anatomical boundary coordinates of the region of interest in the normalized ultrasonic image sequence;

[0008] Perform multi-scale anatomical structure decomposition on each frame of the normalized ultrasound image based on the anatomical boundary coordinates to generate a set of local feature maps corresponding to different tissue layers. Each local feature map in the set of local feature maps contains a three-dimensional matrix of edge sharpness index, texture consistency index, and blood flow signal intensity distribution;

[0009] Input the set of local feature maps into a cascaded deep classification network. Through the spatio-temporal feature alignment module and multi-granularity attention allocation module nested in the cascaded deep classification network, perform cross-frame feature fusion and dynamic weight adjustment on the set of local feature maps, and output the classification probability distribution and spatial topology relationship map of the abnormal region in each frame of the normalized ultrasound image;

[0010] Generate a multi-modal diagnostic report according to the classification probability distribution and spatial topology relationship map, and compare the similarity of the multi-modal diagnostic report with the historical case features in the medical image database to generate an interactive three-dimensional visualization interface including risk level labels and treatment suggestions.

[0011] On the other hand, an embodiment of the present invention also provides an ultrasound image data classification system based on artificial intelligence, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiments of the present application improve the accuracy and efficiency of ultrasonic diagnosis. By real-time acquiring and processing the ultrasonic scan signal sequence, noise is effectively suppressed and motion artifacts are compensated, ensuring the image quality. The multi-scale anatomical structure decomposition technology, combined with the comprehensive analysis of edge sharpness, texture consistency, and blood flow signal intensity, deeply excavates the fine-grained features in the ultrasonic image. The application of the cascaded deep classification network realizes the intelligent fusion of cross-frame features and dynamic weight adjustment, accurately captures the spatio-temporal changes of the abnormal region. The finally generated multi-modal diagnostic report not only provides comprehensive classification probability and spatial topology information, but also outputs intuitive risk levels and treatment suggestions through similarity comparison with historical cases, and constructs a highly interactive three-dimensional visualization interface, greatly enhancing doctors' understanding of the condition and decision-making support. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a schematic execution flow diagram of the method for classifying ultrasound image data based on artificial intelligence provided by the embodiment of the present invention.

[0014] Figure 2 is a schematic diagram of exemplary hardware and software components of the ultrasound image data classification system based on artificial intelligence provided by the embodiment of the present invention. Specific Embodiments

[0015] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 FIG. 5 is a schematic flowchart of a method for classifying ultrasonic image data based on artificial intelligence provided by an embodiment of the present invention. The method for classifying ultrasonic image data based on artificial intelligence will be introduced in detail below.

[0016] Step S110: Obtain a real-time ultrasonic scanning signal sequence of a target object, where the real-time ultrasonic scanning signal sequence includes multiple frames of dynamic ultrasonic images, and the time-sequential change characteristics of tissue elasticity parameters and hemodynamic parameters are embedded in each frame of the dynamic ultrasonic image.

[0017] In this embodiment, taking a patient with liver disease as the target object, a doctor uses an ultrasonic diagnostic device to scan the liver area of the patient. The ultrasonic waves emitted by the ultrasonic device propagate in the liver tissue and are reflected back. After being received by the ultrasonic device, they are converted into electrical signals, thereby forming a real-time ultrasonic scanning signal sequence. Among them, the ultrasonic scanning signal sequence includes multiple frames of dynamic ultrasonic images. For example, in the first frame of the image, the tissue elasticity parameter may reflect that a certain area of the liver has a relatively hard texture, which may imply the presence of fibrosis or sclerosis in this area. The time-sequential change characteristics of the hemodynamic parameter are reflected in that, with the heartbeat cycle, the blood flow velocity and blood flow volume in the blood vessels in the liver show regular changes. For example, during the systolic phase of the heart, the blood flow velocity into the liver increases, and the blood flow volume increases, which is manifested as an increase in the signal intensity in the blood vessels in the image; during the diastolic phase, it is the opposite. Each frame of the image over time continuously records the changes in these tissue elasticity parameters and hemodynamic parameters. For example, in subsequent frames of the image, it may be found that the tissue elasticity around a suspected lesion area in the liver gradually decreases, while the blood flow velocity shows local abnormal fluctuations. These changes are valuable for subsequent judgment of the nature and development trend of the lesion.

[0018] Step S120: Perform noise suppression and motion artifact compensation processing on the real-time ultrasonic scanning signal sequence to generate a denoised and standardized ultrasonic image sequence, and mark the anatomical boundary coordinates of the region of interest in the standardized ultrasonic image sequence.

[0019] Continuing with the example of liver patients, the original real-time ultrasound scan signal sequence has noise due to factors such as the electronic noise of the ultrasound device itself and the interference of the surrounding environment. First, wavelet transform decomposition is performed on each frame of the ultrasound scan signal in the real-time ultrasound scan signal sequence. For example, for a frame of ultrasound scan signal containing the whole liver and some surrounding tissues, after wavelet transform decomposition, the high-frequency noise component may appear as some chaotic, weak-intensity and irregular signal fluctuations, while the low-frequency tissue signal component reflects the general outline of the main structures such as the liver tissue and blood vessels. Then, an adaptive threshold algorithm is used to filter the high-frequency noise component layer by layer. During this process, the set threshold will be dynamically adjusted according to the characteristics of the signal. For example, for areas where the liver edge is relatively blurred, the threshold will be appropriately relaxed to retain weak signals that may belong to the liver tissue, while for obvious noise areas, strict filtering will be performed. After filtering, the intermediate-frequency signal component related to the movement of the blood vessel wall is retained, such as the blood flow signal of the portal vein in the liver.

[0020] Next, inverse wavelet transform reconstruction is performed on the filtered ultrasound scan signal to obtain a preliminarily denoised ultrasound image. However, due to the patient's possible slight breathing or body movement during the scan, motion artifacts exist in the image. Based on the optical flow method, the tissue displacement vector field between adjacent frames of the ultrasound scan signal is calculated. For the liver image, it is to calculate the moving direction and distance of the liver tissue between adjacent frames. According to this tissue displacement vector field, motion trajectory compensation is performed on the preliminarily denoised ultrasound image. For example, a certain feature point on the liver has a certain displacement due to the patient's breathing between adjacent frames, and through compensation, its position in the image becomes more accurate. After that, gray normalization operation is performed on the motion-compensated ultrasound image to map the pixel intensity to a preset standard dynamic range. Assuming the preset standard dynamic range is 0 - 255, the gray values of normal tissues inside the liver may be adjusted to a specific interval, such as 100 - 150, and the diseased area may have different gray value ranges.

[0021] Then, the anisotropic diffusion algorithm is used to perform edge-preserving smoothing on the normalized ultrasound image. During this process, in the edge area of the liver image, the algorithm will reduce the influence of noise while keeping the edge sharp. Then, the bilateral filter is used to enhance the details of the smoothed ultrasound image, such as making the boundaries of blood vessels inside the liver clearer and the levels of blood flow signals inside the blood vessels more distinct. Finally, according to the preset anatomical structure template, regional matching is performed on the detail-enhanced ultrasound image, and a standardized ultrasound image sequence containing the boundary of the complete liver organ is cropped, and the anatomical boundary coordinates of the liver are marked, such as the coordinates of the upper edge, lower edge, and boundaries of the left and right lobes of the liver.

[0022] Step S130: Perform multi-scale anatomical structure decomposition on each frame of the normalized ultrasound image based on the anatomical boundary coordinates, generating a set of local feature maps corresponding to different tissue layers. Each local feature map in the set of local feature maps contains a three-dimensional matrix of edge sharpness index, texture consistency index, and blood flow signal intensity distribution.

[0023] When processing the normalized ultrasound images of liver patients, according to the previously marked liver anatomical boundary coordinates, a rectangular region of interest (ROI) containing the liver is cropped from each frame of the normalized ultrasound image. For example, this rectangular region just covers most of the parenchymal part of the liver and the main vascular branches. Then, Gaussian pyramid decomposition is performed on this rectangular ROI, generating a multi-scale image group containing the original resolution, half resolution, and quarter resolution. For the image with the original resolution, it can display the most delicate structure of the liver tissue, such as tiny vascular branches and the general outline of liver lobules; the image with half resolution generalizes the structural features of a larger area of the liver to a certain extent, which is more beneficial for grasping the overall structure; the image with quarter resolution can highlight the macroscopic features such as the overall shape of the liver and the main vascular trends.

[0024] Apply a multi-directional Gabor filter bank to each image at each scale in the multi-scale image group. Taking the blood vessels in the liver as an example, the Gabor filter bank can extract the frequency domain response feature maps that match the direction of blood vessel branches. In the image with the original resolution, the frequency domain features of tiny blood vessel branches may be accurately extracted, reflecting the texture direction of the blood vessel wall and the local changes in the blood flow signal. The frequency domain response feature maps at different scales are input into a three-dimensional convolutional neural network, and the edge gradient information at different resolutions is aggregated through a cross-scale feature fusion layer. For example, at the edge of the liver, the edge gradient information in the image with the original resolution may be very detailed, while the edge gradient information in the image with quarter resolution focuses more on the overall contour. After fusion, a more comprehensive and accurate liver edge feature can be obtained.

[0025] After that, non-maximum suppression processing is performed on the aggregated feature maps to eliminate redundant edge responses in the liver image and retain the maximum intensity edge points. For example, in the complex vascular network in the liver, some weak edges caused by noise or artifacts are removed, and only the strongest edge points representing important structures such as blood vessel boundaries are retained. A spatial Delaunay triangular mesh is constructed based on the retained edge points, and the curvature distribution and area ratio index of each triangular patch of the triangular mesh are calculated. For the curved parts of the liver surface or internal structure, such as the dome of the liver, the curvature distribution of the triangular patch can reflect the shape change, and the area ratio index can reflect the relative importance of different regions in the overall structure.

[0026] Finally, the triangular mesh is divided into regions according to the curvature distribution and area ratio index, and a set of sub-meshes corresponding to different tissue types is generated. In the liver, different sub-meshes such as the liver parenchyma region and the vascular region may be divided. The spatio-temporal variation variance of the blood flow signal intensity is statistically calculated within each sub-mesh of the sub-mesh set, and a set of local feature maps containing edge geometric features and hemodynamic features is generated. For example, within the sub-mesh of the liver parenchyma region, the spatio-temporal variation variance of the blood flow signal intensity may be relatively small, while within the vascular sub-mesh near the lesion area, due to the influence of the lesion on the blood flow, the spatio-temporal variation variance of the blood flow signal intensity may increase, and these features are recorded in the edge sharpness index, texture consistency index and three-dimensional matrix of the blood flow signal intensity distribution in the set of local feature maps.

[0027] Step S140, input the set of local feature maps into a cascaded deep classification network, and through the spatio-temporal feature alignment module and multi-granularity attention allocation module nested in the cascaded deep classification network, perform cross-frame feature fusion and dynamic weight adjustment on the set of local feature maps, and output the classification probability distribution and spatial topology relationship map of the abnormal region in each frame of the standardized ultrasound image.

[0028] The set of local feature maps of liver patients is input into a cascaded deep classification network. First, it enters the long short-term memory network of the spatio-temporal feature alignment module for temporal state transfer. For example, in a series of consecutive frames of liver ultrasound images, the lesion area may change in morphology, size or blood flow characteristics over time. The long short-term memory network can capture the laws of these changes and generate a time-dimensional feature vector containing the morphological evolution law of the lesion between consecutive frames. For example, the lesion area appears circular and has a relatively clear boundary in the first few frames of images, and gradually becomes irregular and has a blurred boundary over time, and these change features are accurately recorded in the time-dimensional feature vector.

[0029] Then, the time-dimensional feature vector is input into a deformable convolutional layer for spatial deformation modeling. In the ultrasound image of the liver, due to the influence of the respiratory movement of the liver itself and the pulsation of the heart, the liver tissue will undergo certain non-linear deformation. The deformable convolutional layer can output a spatial deformation feature matrix that matches the non-linear deformation of the liver tissue structure, accurately reflecting the actual morphological changes of the liver in space.

[0030] Concatenate the time - dimension feature vectors and the spatial deformation feature matrices to generate a spatio - temporal joint feature matrix, and input it into the multi - head self - attention mechanism of the multi - granularity attention allocation module. For different feature channels in the liver image, such as channels representing liver tissue texture, blood flow signals, edge features, etc., the multi - head self - attention mechanism calculates the global correlation weights between different feature channels and outputs a global attention weight matrix. For example, when judging whether a liver lesion is malignant, the blood flow signal feature channel may have a high correlation with the malignancy degree of the lesion. Then, in the global attention weight matrix, the weight corresponding to the blood flow signal feature channel will be higher.

[0031] Input the global attention weight matrix into a gated linear unit for feature channel screening, retain the feature channels related to the current classification task (judging abnormal areas of the liver), and generate a gated screening feature vector. For example, some feature channels related to the interference of surrounding tissues of the liver may be screened out, while the channels related to the internal structure and lesion characteristics of the liver are mainly retained. Then input the gated screening feature vector into the local window attention calculation layer to enhance the spatial correlation of adjacent feature nodes and generate a local window attention weight matrix. In the liver image, there is a certain spatial correlation between the feature nodes of adjacent liver lobules or blood vessel branches. This multi - granularity attention allocation module can enhance this correlation.

[0032] Weightedly fuse the global attention weight matrix and the local window attention weight matrix to generate a multi - granularity attention allocation map, and stack it to the spatio - temporal joint feature matrix through residual connection to generate an optimized classification feature vector. Finally, input the optimized classification feature vector into a fully - connected classification layer to output the classification probability distribution and the spatial topological relationship map of abnormal areas (such as possible tumors, fibrosis areas, etc.) in each frame of the standardized ultrasound image. The classification probability distribution represents the likelihood of each area of the liver being an abnormal area, and the spatial topological relationship map accurately depicts the spatial relationship between the abnormal area and surrounding normal tissues, blood vessels and other structures, such as whether the abnormal area is close to large blood vessels and whether it invades the surrounding liver tissue, etc.

[0033] Step S150: Generate a multi - modal diagnostic report according to the classification probability distribution and the spatial topological relationship map, and compare the multi - modal diagnostic report with the historical case features in the medical image database to generate an interactive three - dimensional visualization interface containing risk level labels and treatment suggestions.

[0034] In detail, a multimodal diagnostic report can be generated based on the classification probability distribution and spatial topological relationship map of liver patients, and the classification probability distribution can be mapped to the spatial coordinates of the standardized ultrasound image sequence to generate an abnormal area overlay with translucent color coding. For example, in the liver image, the area with a high probability of being abnormal is covered with a red translucent color, the low probability area is covered with a yellow translucent color, and the normal area is kept in the original color or the normal range is indicated by a green translucent color. The spatial topological relationship map is input into the three-dimensional volume rendering engine to construct a three-dimensional volume rendering model with annotated vascular branches and lesion infiltration range. In the three-dimensional model of the liver, the direction of the blood vessels in the liver and the infiltration of abnormal areas (such as tumors) into the surrounding tissues can be clearly seen.

[0035] The abnormal area coverage layer is spatially registered with the 3D volume rendering model to generate a visualization framework that integrates the 2D probability distribution and the 3D anatomical structure. Then, the hemodynamic characteristic parameters in the classification probability distribution are extracted, and the historical case characteristics with similar hemodynamic characteristic parameters are retrieved in the medical imaging database to generate case similarity comparison results. For liver patients, some historical liver disease cases with similar abnormal blood flow velocity and blood flow changes may be found.

[0036] The case similarity comparison results are associated with the visualization basic framework to generate a side comparison view and a key difference parameter annotation layer. For example, in the side comparison view, the liver images, hemodynamic parameters and other information of the current patient and historical cases can be displayed simultaneously, and the key difference parameter annotation layer will highlight the differences between the current patient and historical cases in terms of liver size, lesion location, blood flow velocity, etc. Based on the key difference parameter annotation layer, the preset treatment plan flow chart is matched to generate a logical association table of risk level labels and treatment recommendations. The logical association table is then embedded in the visualization basic framework to generate an interactive three-dimensional visualization interface that includes an interactive timeline control, real-time ultrasound playback function and voice broadcast engine. Doctors can view liver ultrasound images and analysis results at different time points through the timeline control, and reconfirm the details in the image through the real-time ultrasound playback function. The voice broadcast engine can broadcast information such as diagnosis results, risk levels and treatment recommendations, which is convenient for doctors to operate during the diagnosis process and make comprehensive judgments on the patient's condition.

[0037] Based on the above steps, the embodiments of the present application improve the accuracy and efficiency of ultrasonic diagnosis. By acquiring and processing the ultrasonic scan signal sequence in real time, noise is effectively suppressed and motion artifacts are compensated, ensuring image quality. The multi-scale anatomical structure decomposition technology, combined with the comprehensive analysis of edge sharpness, texture consistency, and blood flow signal intensity, deeply explores the fine-grained features in ultrasonic images. The application of the cascaded deep classification network realizes the intelligent fusion of cross-frame features and dynamic weight adjustment, accurately captures the spatio-temporal changes in abnormal regions, and finally generates a multi-modal diagnostic report that not only provides comprehensive classification probabilities and spatial topology information but also outputs intuitive risk levels and treatment suggestions through similarity comparison with historical cases, and constructs a highly interactive three-dimensional visualization interface, greatly enhancing doctors' understanding of the condition and decision-making support.

[0038] In a possible implementation manner, step S120 includes:

[0039] Step S121, perform wavelet transform decomposition on each frame of the real-time ultrasonic scan signal sequence to separate the high-frequency noise component and the low-frequency tissue signal component.

[0040] Taking the ultrasonic scan of a previous liver patient as an example, in this process, it is assumed that the wavelet transform uses the Daubechies wavelet basis function (such as Db4 wavelet), which can effectively decompose the signal into sub-bands of different frequencies. For a frame of liver ultrasonic scan signal, after wavelet transform decomposition, the high-frequency noise component usually appears as the signal part with relatively high frequency, relatively low energy, and no obvious pattern. These high-frequency noises may originate from the electronic components of the ultrasonic device, minute interferences in the surrounding environment, etc. The low-frequency tissue signal component contains the main structural information of the liver tissue, such as the overall contour of the liver and the general trend of larger blood vessels, with relatively low frequency and concentrated energy.

[0041] Step S122, use an adaptive threshold algorithm to filter the high-frequency noise component layer by layer to obtain the filtered ultrasonic scan signal, and retain the intermediate-frequency signal component related to the movement of the blood vessel wall.

[0042] For example, for the first layer of the high-frequency noise component of a certain frame of liver ultrasound scan signal, the algorithm analyzes statistical parameters such as the standard deviation and mean of the signal of this layer. If the standard deviation is 5 (assuming the unit), then according to a certain algorithm rule (such as multiplying the standard deviation by a coefficient, assuming the coefficient is 3), the threshold of this layer is obtained as 15. The signal part higher than this threshold is regarded as a potentially useful signal and retained, while the part lower than the threshold is determined as noise and filtered out. During the entire layer-by-layer filtering process, this adaptive threshold adjustment can effectively remove noise while retaining the intermediate-frequency signal components related to the movement of the blood vessel wall. For the blood vessels in the liver, the signal frequency generated by their blood flow is in the intermediate-frequency range, for example, between 1 - 5 MHz (assuming), and these intermediate-frequency signals are of great significance for subsequent analysis of the movement state of the blood vessel wall, blood flow velocity, etc.

[0043] Step S123, perform inverse wavelet transform reconstruction on the filtered ultrasound scan signal to generate a preliminarily denoised ultrasound image.

[0044] After completing the filtering of the high-frequency noise component, perform inverse wavelet transform reconstruction on the filtered ultrasound scan signal. This operation recombines the sub-band signals after filtering according to the inverse process of wavelet transform to generate a preliminarily denoised ultrasound image. In this preliminarily denoised liver ultrasound image, some blurred parts of the liver tissue become relatively clear due to the removal of high-frequency noise, and the outline of the blood vessels is also more obvious. However, due to factors such as the patient's breathing and slight body movement during the scan, there may be motion artifacts in the image.

[0045] Step S124, calculate the tissue displacement vector field between adjacent frames of ultrasound scan signals based on the optical flow method, and perform motion trajectory compensation on the preliminarily denoised ultrasound image according to the tissue displacement vector field. Perform gray normalization on the motion-compensated ultrasound image to map the pixel intensity to a preset standard dynamic range to obtain a normalized ultrasound image.

[0046] To eliminate motion artifacts, the tissue displacement vector field between adjacent frame ultrasound scan signals is calculated based on the optical flow method. The optical flow method determines the motion direction and speed of an object by analyzing the brightness changes of pixels in adjacent frame images. For the ultrasound images of liver patients, assuming the time interval between two adjacent frames is 0.1 seconds (a typical time interval determined by the scan frame rate of the ultrasound device), the optical flow method calculates the displacement vector of the liver tissue within this 0.1 second. For example, if a specific point (assuming the coordinates are (x1, y1)) on the liver moves to (x2, y2) in the second frame image from its position in the first frame image, the calculated displacement vector is ((x2 - x1), (y2 - y1)). The motion trajectory compensation is performed on the preliminarily denoised ultrasound image according to this tissue displacement vector field. If a suspected lesion area in the liver has an obvious displacement in the image due to the patient's breathing movement before compensation, after the motion trajectory compensation, the position of the suspected lesion area in the image can more accurately reflect its true position in the liver.

[0047] After that, a gray normalization operation is performed on the motion-compensated ultrasound image. Assuming the preset standard dynamic range is 0 - 255 (this is a common image pixel intensity range), in the motion-compensated liver ultrasound image, first, the gray value range of all pixels in the image is statistically analyzed. For example, the gray value range of the liver tissue may be 50 - 180, and the gray value range of blood vessels may be 30 - 100 (these are just assumed approximate ranges here). Then, through linear mapping, these original gray values are mapped to the standard dynamic range of 0 - 255. The purpose of this is to make the images of different frames and different patients comparable in terms of gray level, facilitating subsequent analysis and processing. After the gray normalization operation, the normalized ultrasound image is obtained.

[0048] In step S125, the anisotropic diffusion algorithm is used to perform edge-preserving smoothing on the normalized ultrasound image, and after detail enhancement of the smoothed ultrasound image through a bilateral filter, region matching is performed on the detail-enhanced ultrasound image according to a preset anatomical structure template, and a standardized ultrasound image sequence containing the complete organ boundary is cropped.

[0049] In this embodiment, when the anisotropic diffusion algorithm processes liver ultrasound images, it adjusts the direction and intensity of diffusion according to the gradient information of the ultrasound images. At the liver edge, due to the large gray gradient of the edge, the algorithm restricts the diffusion perpendicular to the edge direction, thus maintaining the sharpness of the edge. For example, for a pixel point on the liver edge, the gray difference between its surrounding pixels is large. During the diffusion process of this pixel point, the algorithm will preferentially smooth along the edge direction to avoid blurring the edge. After being processed by the anisotropic diffusion algorithm, the smoothed ultrasound image is further enhanced in detail by a bilateral filter. The bilateral filter takes into account the spatial distance and gray difference of pixels. For structures such as blood vessels and hepatic lobules inside the liver, it can enhance their detailed features. For example, after the blood flow signal inside the blood vessel is processed by the bilateral filter, the signal has stronger layering and the boundary of the blood vessel wall is clearer.

[0050] The preset liver anatomical structure template contains information such as the standard shape, size, and positions of the main blood vessel branches of the liver. For the liver ultrasound image enhanced in detail, the liver region corresponding to the template in the image is found through a template matching algorithm. For example, the template stipulates that the coordinate range of the upper edge of the liver in the image is (y1 - y2), and the algorithm finds the region in the image that meets this coordinate range and has the characteristics of liver tissue. Finally, a standardized ultrasound image sequence containing the boundaries of the complete liver organ is cropped. Each frame of the image in this standardized ultrasound image sequence has undergone the above series of processes, has low noise, clear structures, and accurate liver region positioning, providing high-quality image data for subsequent analysis.

[0051] In a possible implementation manner, step S130 includes:

[0052] Step S131, in each frame of the standardized ultrasound image, intercept a rectangular region of interest containing the target organ according to the anatomical boundary coordinates.

[0053] In this embodiment, the anatomical boundary coordinates precisely define the position of the liver in the image. For example, the coordinate range of the upper edge of the liver is from the 100th to the 150th row of pixels from the top of the image (assuming the total number of rows of the image is 500), and the coordinate ranges of the left and right edges are from the 200th to the 300th column of pixels and from the 400th to the 500th column of pixels from the left of the image respectively (assuming the total number of columns of the image is 600). The rectangular region of interest determined by these coordinates covers most of the parenchymal part of the liver and the main blood vessel branches, so that subsequent analysis can focus on the liver region and avoid interference from surrounding tissues.

[0054] Step S132, perform Gaussian pyramid decomposition on the rectangular region of interest to generate a multi-scale image group including the original resolution, half resolution, and quarter resolution.

[0055] For the image of the original resolution, it preserves the finest structural information of the liver region. For example, the tiny blood vessel branches within the liver, whose diameters may be as small as 1 - 2 pixels (assuming that one pixel in the image represents an actual length of 0.1 mm), and the general outline of the hepatic lobules can also be clearly seen. The image of half resolution is obtained by downsampling the image of the original resolution. At this scale, the structural features of larger regions of the liver are more obvious. For example, the features of larger structures such as the interlobar fissures of the liver are easier to grasp as a whole, and some tiny blood vessel branches may be merged into thicker blood vessels for representation, but the overall blood vessel trend is still clear. The image of quarter resolution further generalizes the macroscopic features of the liver, and the trend of the main blood vessel branches is more prominent, and the overall shape of the liver and the general distribution of the tissue structure can be seen at a glance.

[0056] Step S133, apply a multi - direction Gabor filter bank to the images of each scale in the multi - scale image group, and extract the frequency - domain response feature maps that match the direction of blood vessel branches.

[0057] For the blood vessels in the liver, due to the certain directionality and texture features of the blood vessels, the Gabor filter bank can capture these features well. For example, assume that the directions of the Gabor filter bank are set to four directions: 0°, 45°, 90°, and 135° (the number of directions and angles can be adjusted according to actual needs). In the liver image of the original resolution, when the Gabor filter is filtered along an angle close to the direction of the blood vessel branch, a strong response can be obtained in the frequency domain. For smaller - diameter blood vessel branches, such as intrahepatic arterioles (assuming a diameter of 2 - 3 pixels), the frequency - domain response feature map will show a specific frequency and amplitude distribution in the direction of the Gabor filter consistent with the blood vessel direction, reflecting the texture direction of the blood vessel wall and the local changes in the blood flow signal. In the images of half resolution and quarter resolution, the same principle applies. Only due to the reduced resolution, the representation of the blood vessel branches is more generalized, and the frequency - domain response feature maps correspondingly reflect the comprehensive features of a larger range of blood vessels.

[0058] Step S134, input the frequency - domain response feature maps of different scales into a three - dimensional convolutional neural network, aggregate the edge gradient information at different resolutions through a cross - scale feature fusion layer, and perform non - maximum suppression processing on the aggregated feature maps. After eliminating redundant edge responses and retaining the maximum - intensity edge points, construct a spatial Delaunay triangular mesh based on the retained edge points, and calculate the curvature distribution and area proportion index of each triangular patch of the triangular mesh.

[0059] For the edge of the liver, the edge gradient information in the original resolution image is very detailed. For example, at the boundary between the liver edge and the surrounding tissues, the gray value of pixels may have a large gradient change within the range of several pixels, such as suddenly changing from 100 (gray value of liver tissue) to 50 (gray value of surrounding tissues). The edge gradient information of the liver in the quarter-resolution image focuses more on the overall contour. Although the details are not as good as those in the original resolution image, it can reflect the changing trend of the liver edge at a larger scale. Through the cross-scale feature fusion layer, the edge gradient information at these different resolutions is fused to obtain a more comprehensive and accurate representation of the liver edge features.

[0060] Furthermore, in the liver image, due to the existence of edge information of various structures, there may be some redundant edge responses. For example, near the complex vascular network and hepatic lobule structure inside the liver, some edge responses generated by noise or weak structures do not truly represent the edges of important structures. Non-maximum suppression processing eliminates those edge responses that are not local maxima by comparing the edge intensity of each pixel point with that of its neighboring pixel points. For the edges of liver blood vessels, only the pixel points that truly represent the blood vessel boundaries (with the maximum edge intensity) will be retained, and other weaker edge responses will be suppressed. In this process, the size of the comparison neighborhood can be set, such as a 3×3 pixel neighborhood (which can be adjusted according to the actual situation), and the edge intensity is compared within this neighborhood.

[0061] Furthermore, for the surface or internal structure of the liver, these edge points form the vertices of a triangular mesh. In curved parts such as the hepatic dome, the construction of the triangular mesh can accurately reflect the shape change. Then, the curvature distribution and area ratio index of each triangular patch of the triangular mesh are calculated. Suppose a triangular patch on the liver is composed of three vertex coordinates (x1, y1, z1), (x2, y2, z2), and (x3, y3, z3) (assuming three-dimensional image coordinates here, and the z-axis represents the depth direction). The curvature distribution is determined by calculating geometric parameters such as the side lengths and angles of the triangle. For example, if the area where the triangular patch is located is relatively flat, its curvature value may be small, such as 0.1 (assuming the unit of curvature); if it is in the curved part of the liver, the curvature value may be large, such as 0.5. The area ratio index is obtained by calculating the ratio of the area of each triangular patch to the total area of the entire triangular mesh. For example, if the area of a certain triangular patch is 10 square pixels (assuming), and the total area of the entire triangular mesh is 100 square pixels, then its area ratio is 0.1.

[0062] Step S135: Divide the triangular mesh according to the curvature distribution and area ratio index to generate a set of sub-meshes corresponding to different tissue types, and calculate the spatio-temporal variance of the blood flow signal intensity in each sub-mesh of the sub-mesh set to generate a set of local feature maps containing edge geometric features and hemodynamic features.

[0063] In the liver, different sub-meshes such as the liver parenchyma region and the blood vessel region can be divided according to these indexes. For example, a region with a smaller curvature and a larger area ratio may be divided into a liver parenchyma sub-mesh, while a region with a larger curvature and related to the blood vessel direction is divided into a blood vessel sub-mesh. Calculate the spatio-temporal variance of the blood flow signal intensity in each sub-mesh of the sub-mesh set. In the liver parenchyma sub-mesh, the blood flow is relatively slow and stable. Assuming that within a certain period of time (such as 5 seconds, assuming the continuous scanning time of the ultrasonic device), the average value of the blood flow signal intensity is 80 (assuming intensity unit), and its fluctuation range is small, the spatio-temporal variance may be 5 (assuming variance unit). In the blood vessel sub-mesh near the lesion area, due to the influence of the lesion on the blood flow, the fluctuation of the blood flow signal intensity is large. Assuming that within the same 5 seconds, the average value of the blood flow signal intensity is 100, but the fluctuation range is large, and the spatio-temporal variance may be 20. Through this statistics, a set of local feature maps containing edge geometric features and hemodynamic features is generated.

[0064] In a possible implementation manner, step S140 includes:

[0065] Step S141: Input the set of local feature maps into the long short-term memory network of the spatio-temporal feature alignment module for temporal state transfer to generate a time-dimensional feature vector containing the morphological evolution law of the lesion between consecutive frames, and input the time-dimensional feature vector into the deformable convolutional layer for spatial deformation modeling to output a spatial deformation feature matrix matching the non-linear deformation of the tissue structure.

[0066] In this embodiment, first, the set of local feature maps of liver patients is input into the long short-term memory network (LSTM) of the spatio-temporal feature alignment module for temporal state transfer. In the ultrasound image sequence of the liver, consecutive multiple frames of images contain information about the changes of liver lesions (if any) over time. Suppose the ultrasound image sequence contains 10 frames of images (the number of frames here is determined according to the actual scanning duration and the device frame rate), and each frame of image contains the previously generated local feature maps, which contain information such as the edge geometric features and hemodynamic features of the liver. The LSTM network gradually learns the morphological evolution law of lesions between consecutive frames through temporal state transfer of the local feature maps corresponding to these 10 frames of images. For example, in the first frame of image, the suspected lesion area in the liver presents an approximate circle with a diameter of about 5 pixels (assuming that one pixel in the image represents an actual length of 0.1 mm). As time progresses to the 5th frame of image, the shape of the lesion area gradually changes to an ellipse, with the long axis direction along the direction of a certain blood vessel in the liver, the long axis length becoming 8 pixels, and the short axis length being 4 pixels. The change information of the lesion in shape, size, and relative position is encoded into the time-dimensional feature vector. This time-dimensional feature vector contains information in multiple dimensions, and each dimension corresponds to different feature parameters, such as shape change parameters, position movement parameters, and change parameters of the relationship with surrounding tissues. The specific values of these parameters are obtained through the calculation of the LSTM network according to the actual changes of the lesion area in the liver image.

[0067] Next, due to the physiological movements of the liver itself (such as the displacement and deformation of the liver caused by respiratory movement and cardiac pulsation) and the influence of possible lesions on surrounding tissues, the liver tissue structure will undergo non-linear deformation. The deformable convolutional layer can model this non-linear deformation of the liver tissue structure according to the information in the time-dimensional feature vector. For example, in a certain area of the liver, which is normally a relatively flat hepatic parenchyma part, but due to being close to a growing lesion tissue, the hepatic parenchyma in this area is compressed and sunken. The deformable convolutional layer outputs a spatial deformation feature matrix that matches this non-linear deformation of the tissue structure by analyzing the relevant parameters in the time-dimensional feature vector. The element values in this spatial deformation feature matrix reflect the degree and direction of deformation of the liver tissue at different positions. For example, in the above-mentioned sunken area, the corresponding matrix element may be expressed as having a displacement of -2 pixels in the x direction (assuming one direction of the image coordinate system), a displacement of -1 pixel in the y direction, and a displacement of -0.5 pixels in the depth z direction (if it is a three-dimensional image). The displacement values here are obtained through the calculation of the deformable convolutional layer according to the actual deformation situation.

[0068] Step S142: Concatenate the time - dimension feature vector and the spatial deformation feature matrix to generate a spatio - temporal joint feature matrix, and input the spatio - temporal joint feature matrix into the multi - head self - attention mechanism of the multi - granularity attention allocation module to calculate the global correlation weights between different feature channels, and output the global attention weight matrix.

[0069] In this embodiment, the spatio - temporal joint feature matrix synthesizes the feature information of liver lesions in time and space. In the feature representation of liver ultrasound images, there are multiple feature channels, such as the channel representing the texture features of liver tissue, the blood flow signal feature channel, the edge feature channel, etc. The multi - head self - attention mechanism calculates the global correlation weights between different feature channels. For judging the abnormal area of the liver, the importance of different feature channels is different. For example, when judging whether a liver lesion is malignant, the blood flow signal feature channel may have a higher correlation with the malignancy degree of the lesion, because malignant lesions are often accompanied by abnormal blood supply. Suppose the correlation weight of the blood flow signal feature channel calculated in this process with lesion judgment is 0.6 (the weight value here is calculated by the multi - head self - attention mechanism according to the data features), while the correlation weight of the tissue texture feature channel is 0.3, and the correlation weight of the edge feature channel is 0.1. Through such calculation, the global attention weight matrix is output, and each element in the global attention weight matrix represents the global correlation weight of the corresponding feature channel.

[0070] Step S143: Input the global attention weight matrix into the gated linear unit for feature channel screening, retain the feature channels related to the current classification task, generate a gated screening feature vector, and input the gated screening feature vector into the local window attention calculation layer to enhance the spatial correlation of adjacent feature nodes, and generate a local window attention weight matrix.

[0071] In this embodiment, the gated linear unit retains the feature channels related to the current classification task (judging the abnormal area of the liver) according to the weight values in the global attention weight matrix. For example, for the feature channels with weight values lower than 0.2 (such as the above - mentioned edge feature channel), the gated linear unit may filter them out, and mainly retain the blood flow signal feature channel and the tissue texture feature channel with higher weight values. After screening, a gated screening feature vector is generated. This gated screening feature vector mainly contains the most valuable feature information for judging the abnormal area of the liver.

[0072] Next, in the liver image, there is a certain spatial correlation among the characteristic nodes between adjacent hepatic lobules or vascular branches. The local window attention calculation layer enhances the spatial correlation of adjacent characteristic nodes by analyzing the characteristic information in the gated screening feature vector. For example, for two adjacent hepatic lobule regions in the liver, their tissue texture features and blood flow signal features have a certain continuity and correlation in space. The local window attention calculation layer generates a local window attention weight matrix by calculating the correlation weight between the characteristic nodes corresponding to these two regions. The element value in this local window attention weight matrix reflects the intensity of the spatial correlation between adjacent characteristic nodes. For example, for two highly correlated hepatic lobule regions, the correlation weight between them may be 0.8 (the weight value here is calculated by the local window attention calculation layer according to the actual characteristic correlation).

[0073] Step S144: Weightedly fuse the global attention weight matrix and the local window attention weight matrix to generate a multi-granularity attention allocation map, and superimpose the multi-granularity attention allocation map on the spatio-temporal joint feature matrix through residual connection to generate an optimized classification feature vector.

[0074] In this embodiment, the multi-granularity attention allocation map synthesizes attention information of different granularities (global and local). For example, for a certain feature channel, its weight in the global attention weight matrix is 0.5, and its weight in the local window attention weight matrix is 0.4. Then, in the multi-granularity attention allocation map, the final weight of this feature channel may be calculated according to a certain weighting rule, assumed to be 0.45 (the weighting rule here can be simple linear weighting or other rules defined according to the network structure). Superimpose this multi-granularity attention allocation map on the spatio-temporal joint feature matrix through residual connection to generate an optimized classification feature vector. Residual connection helps to retain part of the information in the original spatio-temporal joint feature matrix, avoid information loss in the multi-layer network processing, and the optimized classification feature vector synthesizes multi-faceted attention information and the original spatio-temporal features.

[0075] Step S145: Input the optimized classification feature vector into the fully connected classification layer to output the classification probability distribution and the spatial topology relationship map of the abnormal region in each frame of the normalized ultrasound image.

[0076] In the scenario of liver ultrasound images, output the classification probability distribution and the spatial topological relationship map of the abnormal regions (such as possible tumors, fibrotic regions, etc. in the liver) in each frame of the standardized ultrasound image. The classification probability distribution represents the likelihood of each region of the liver being an abnormal region. For example, if the probability of a certain region in the liver being calculated as an abnormal region is 0.8, it means that according to a series of previous feature analyses and model calculations, there is an 80% likelihood that this region is abnormal. The spatial topological relationship map accurately depicts the spatial relationships between the abnormal regions and surrounding normal tissues, blood vessels and other structures. For example, whether the abnormal region is close to large blood vessels, whether the boundary with the surrounding liver tissue is clear or blurred, and whether it has invaded the surrounding hepatic lobules, etc. These information are of great significance for accurately diagnosing liver diseases, evaluating the severity of lesions, and formulating subsequent treatment plans.

[0077] In a possible implementation manner, step S150 includes:

[0078] Step S151, map the classification probability distribution to the spatial coordinates of the standardized ultrasound image sequence to generate an abnormal region overlay with semi-transparent color coding.

[0079] In this embodiment, for the diagnosis process of liver patients, first, map the classification probability distribution of liver patients to the spatial coordinates of the standardized ultrasound image sequence to generate an abnormal region overlay with semi-transparent color coding. In the standardized ultrasound image sequence of the liver, each pixel point has its corresponding spatial coordinates. The classification probability distribution represents the likelihood of each region of the liver being an abnormal region. For example, for a pixel point in the liver image with coordinates (x, y) (assuming the image coordinate system), if the probability of this point being classified as an abnormal region is 0.8 (this probability value is obtained according to previous classification calculations), then this pixel point is encoded and marked with a specific semi-transparent color. For regions with a high probability of being abnormal, such as regions with a probability greater than 0.7, they are covered with a red semi-transparent color, indicating highly suspected lesion regions; for regions with a probability between 0.4 - 0.7, they are covered with a yellow semi-transparent color, representing regions with possible lesion risks; for regions with a probability less than 0.4, they are covered with a green semi-transparent color, indicating normal or low-risk regions. In this way, the entire liver image is covered with a layer of abnormal region overlay with color coding, and doctors can intuitively see the lesion risk distribution of different regions of the liver.

[0080] Step S152, input the spatial topological relationship map into a three-dimensional volume rendering engine to construct a three-dimensional volume rendering model that annotates the blood vessel branches and the infiltration range of the lesion.

[0081] For example, the spatial topological relationship map of the liver contains the spatial relationship information between various structures in the liver (such as blood vessels, liver lobules, lesion areas, etc.). The 3D rendering engine constructs a 3D model of the liver based on this information. In this 3D model, the vascular branches in the liver are clearly marked, and information such as the direction and diameter of the portal vein and its branches can be accurately displayed. For the range of lesion infiltration, if there is a tumor lesion in the liver, the model will accurately show the location and size of the tumor in the liver and the infiltration relationship with the surrounding tissues. Assume that the tumor in the liver is located in the right lobe of the liver, its center coordinates are (x1, y1, z1) (here the z axis represents the depth direction), and the diameter is 10 pixels (assuming that one pixel represents an actual length of 0.1 mm). The boundary between the tumor periphery and the surrounding liver tissue is represented in the model with different colors or textures to show its degree of infiltration.

[0082] Step S153: spatially aligning the abnormal region cover layer with the three-dimensional volume rendering model to generate a visualization basic framework that integrates the two-dimensional probability distribution and the three-dimensional anatomical structure.

[0083] In this process, it is necessary to ensure that each pixel in the abnormal area coverage layer accurately matches the corresponding position in the 3D volume rendering model in space. For example, the high-risk area marked with red semi-transparent marks in the abnormal area coverage layer must also be accurately located in the corresponding position in the liver in the 3D volume rendering model. Through this spatial registration, the generated visualization basic framework combines the two-dimensional probability distribution information (abnormal area coverage layer) with the three-dimensional anatomical structure information (3D volume rendering model). Doctors can view the two-dimensional risk distribution and three-dimensional structural relationship of the liver in one view, and have a more comprehensive understanding of the health of the liver.

[0084] Step S154, extracting the hemodynamic characteristic parameters in the classification probability distribution, searching the medical image database for historical case characteristics with similar hemodynamic characteristic parameters, and generating case similarity comparison results.

[0085] In the classification probability distribution of liver patients, hemodynamic characteristic parameters include information such as blood flow velocity, blood flow volume, and blood flow direction of blood vessels in the liver. For example, the blood flow velocity of a major blood vessel in the liver is 50 cm / s (this value is obtained based on previous ultrasound image analysis), and the blood flow volume is 100 ml / min. In the medical imaging database, a large amount of image data of historical cases of liver diseases and their related characteristics are stored. By retrieving the hemodynamic characteristic parameters of cases in the database, historical cases with similar hemodynamic characteristics to the current patient are found. Suppose 3 historical cases are found, and the differences in their hemodynamic characteristic parameters from the current patient are within a certain allowable range (for example, the difference in blood flow velocity is within ±10%, and the difference in blood flow volume is within ±15%). Then, a detailed comparison is made between these historical cases and the current patient, including aspects such as the size of the liver, the location of the lesion, and the type of the lesion, to generate a case similarity comparison result.

[0086] Step S155: Associate the case similarity comparison result with the visualization basic framework to generate a side-by-side comparison view and a key difference parameter annotation layer, match a preset treatment plan flowchart based on the key difference parameter annotation layer to generate a logical association table of risk level labels and treatment suggestions, and then embed the logical association table into the visualization basic framework to generate an interactive three-dimensional visualization interface including an interactive timeline control, a real-time ultrasound playback function, and a voice broadcast engine.

[0087] In the side-by-side comparison view, relevant information of the current patient and historical cases is displayed simultaneously. For example, on one side, a three-dimensional model of the current patient's liver with an abnormal area overlay is displayed, and information such as its hemodynamic characteristic parameters and liver size is also displayed; on the other side, the same type of information of the historical cases is displayed. The key difference parameter annotation layer will highlight the important differences between the current patient and historical cases. For example, if the size of the liver tumor of the current patient is 20 cubic millimeters, while the tumor size in the historical case is 15 cubic millimeters, then this size difference will be clearly marked in the key difference parameter annotation layer. Based on the key difference parameter annotation layer, a preset treatment plan flowchart is matched to generate a logical association table of risk level labels and treatment suggestions. The preset treatment plan flowchart includes different treatment plans and corresponding risk levels formulated according to different liver disease characteristics (such as lesion size, location, hemodynamic characteristics, etc.). For example, if the size of the liver tumor is less than 10 cubic millimeters and the hemodynamic characteristics are normal, the risk level may be low risk, and the corresponding treatment suggestion may be regular observation; if the tumor size is between 10 - 20 cubic millimeters, the risk level is medium risk, and the treatment suggestion may be drug treatment or local interventional treatment. According to the differences between the current patient and historical cases and these preset rules, a logical association table of risk level labels and treatment suggestions is generated.

[0088] Finally, the interactive timeline control allows doctors to view liver ultrasound images and related analysis results at different time points. For example, doctors can use the timeline to go back to a previous ultrasound examination and view the liver's status and analysis results at that time. The real-time ultrasound playback function enables doctors to watch the real-time process of the liver ultrasound scan again, so as to observe certain suspicious areas more carefully. The voice announcement engine can announce information such as diagnosis results, risk levels, and treatment suggestions, facilitating doctors to obtain information during the operation. For example, the voice announcement will clearly state content such as "There is a suspected tumor lesion in the right lobe of the liver of the current patient, the risk level is medium risk, and the treatment suggestions are drug treatment or local interventional treatment". Such an interactive three-dimensional visualization interface provides doctors with a comprehensive and convenient diagnostic tool, helping to improve the diagnostic accuracy of liver diseases and the rationality of treatment decisions.

[0089] In a possible implementation manner, the method further includes:

[0090] Step S210, input the training sample library with multi-dimensional labels into the data augmentation module, perform random elastic deformation, local pixel occlusion, and multi-angle affine transformation to generate an augmented adversarial training sample set.

[0091] Specifically, the training sample library contains numerous liver ultrasound image samples, and each sample is accompanied by multi-dimensional labels. These labels include information such as the liver's health status (normal or lesion type), hemodynamic characteristic parameters, and tissue elasticity parameters. For the random elastic deformation operation, taking the liver region in the liver ultrasound image as an example, within the coordinate space of the image, the shape of the liver is changed according to certain elastic deformation rules. Assume that the size of the liver image is 500×600 pixels (the pixel size here is determined according to the imaging resolution of the ultrasound device). During the random elastic deformation process, some pixel points on the liver edge may generate displacements in the x direction (horizontal direction) and y direction (vertical direction). For example, a pixel point on the upper edge of the liver, originally with coordinates (200, 100), may become (203, 98) after elastic deformation. This deformation simulates the image deformation situation caused by the patient's breathing, body movement, or the liver's own physiological movements (such as the slight displacement of the liver caused by the heartbeat) during the actual ultrasound scan.

[0092] Local pixel occlusion means randomly selecting some pixels in small regions of the image for occlusion. For liver images, some small regions inside the liver may be selected, such as a 10×10 pixel region in the right lobe of the liver (the size of this region is determined according to the preset occlusion rules) for occlusion. This simulates the situation where some tissues may be occluded by other structures during actual scanning or local interference during the image acquisition process. Multi-angle affine transformation includes operations such as rotating, scaling, and translating the image. Taking rotation as an example, the liver ultrasound image is rotated around the center of the image by a certain angle, such as rotating 15° (this angle is a randomly determined affine transformation parameter). Through these random elastic deformations, local pixel occlusions, and multi-angle affine transformation operations, the samples in the original training sample library are expanded to generate an expanded adversarial training sample set. The number of samples in this adversarial training sample set has increased significantly compared to the original training sample library, and it contains more diverse sample forms, which helps to improve the generalization ability of the subsequent network.

[0093] Step S220: Input the expanded adversarial training sample set into the focal loss function in the first training stage to optimize the inter-class balance weight parameters of the cascaded deep classification network.

[0094] In this cascaded deep classification network, there are different classes, such as normal liver, liver fibrosis, liver tumor, etc. The focal loss function aims to solve the problem of unbalanced sample numbers of different classes during training. Suppose in the expanded adversarial training sample set, there are 500 normal liver samples, 300 liver fibrosis samples, and 200 liver tumor samples. Due to the difference in sample numbers, under the traditional loss function, the network may tend to classify the more numerous normal samples better while ignoring the less numerous fibrosis and tumor samples. The focal loss function adjusts the losses of different class samples by weighting, making the network pay more attention to those minority class samples that are difficult to classify. During training, the loss value of each sample is calculated according to the focal loss function, and the inter-class balance weight parameters of the cascaded deep classification network are updated based on these loss values. For example, for liver tumor samples, since their number is small and they are difficult to classify, the focal loss function will increase the weight of their losses, so that the network pays more attention to learning the features of liver tumor samples during the learning process, adjusts the inter-class balance weight parameters, and improves the classification accuracy of different class samples.

[0095] Step S230: Input the optimized inter-class balance weight parameters into the adversarial training strategy in the second training stage to generate an adversarial perturbation sample set through gradient sign attack.

[0096] In the adversarial training strategy, based on the optimized inter-class balance weight parameters, an adversarial perturbation sample set is generated by means of gradient sign attack. For a cascaded deep classification network, it is assumed that the predicted result is obtained through the forward propagation of the network for the input liver ultrasound image sample. During the gradient sign attack process, the gradient of the loss function with respect to the input sample is calculated according to the predicted result and the true label. Taking a liver ultrasound image sample as an example, its image data is a matrix of 500×600×3 (width×height×number of channels, where the number of channels is 3 indicating a color image). The calculated gradient represents the rate of change of the loss function as the pixel values of the input image change. Then, the adversarial perturbation is generated according to the sign of the gradient (positive or negative). For example, at a certain pixel point of the image, if the gradient is positive, the value of that pixel point is increased; if the gradient is negative, the value of that pixel point is decreased. The samples in the training sample set are perturbed in this way to generate an adversarial perturbation sample set. These adversarial perturbation samples can increase the robustness of the network to adversarial samples, that is, improve the correct classification ability of the network when facing input samples that may be maliciously modified or have interference.

[0097] Step S240: Input the adversarial perturbation sample set into the knowledge distillation framework in the third training stage, and use the soft label transfer mechanism of the teacher network to compress the model complexity parameters of the student network.

[0098] In the knowledge distillation framework, there are a teacher network and a student network. The teacher network is a pre-trained network with high accuracy and complexity. It makes predictions on the adversarial perturbation sample set, and the obtained prediction results exist in the form of soft labels. The soft labels not only contain the probability information of the sample belonging to each class, but also contain more uncertainty information. For example, for an adversarial perturbation sample of a liver ultrasound image, the teacher network predicts that the probability of it being normal liver is 0.6, the probability of liver fibrosis is 0.3, and the probability of liver tumor is 0.1. The student network is a relatively simple network with low model complexity. It tries to learn useful information from the soft labels of the teacher network. Through this soft label transfer mechanism, the student network does not need to re-learn complex feature representations, but obtains information such as the relationships between classes from the soft labels of the teacher network, thereby compressing its own model complexity parameters while maintaining a certain degree of accuracy. For example, the student network may originally have more convolutional layers and neuron connections. After knowledge distillation, some unnecessary convolutional layers or neuron connections are removed, thus reducing the model complexity and improving the training efficiency and generalization ability of the network.

[0099] Step S250: Input the compressed model complexity parameters into the validation set monitoring module, and trigger the early stopping mechanism according to the classification accuracy and generalization performance and roll back to the optimal weight snapshot.

[0100] The validation set contains a part of liver ultrasound image samples and their labels that are independent of the training set. The network corresponding to the compressed model complexity parameters is applied to the validation set for testing. During this process, the classification accuracy of the network on the validation set is calculated, for example, calculating the proportion of the number of correctly predicted samples to the total number of validation samples. At the same time, the generalization performance of the network is evaluated. The generalization performance can be measured by comparing the performance on the validation set and the performance on the training set. If the performance on the validation set is not much different from the performance on the training set, it indicates that the network has good generalization ability. Suppose the validation set has 200 samples. When the classification accuracy reaches 90% (this value is a performance index set according to actual requirements) and the generalization performance is good, the validation set monitoring module will consider that the current network has reached a good state. If during the subsequent training process, the classification accuracy starts to decline or the generalization performance deteriorates, the early stopping mechanism will be triggered, and the network will roll back to the weight snapshot when the optimal classification accuracy and generalization performance were achieved before. This weight snapshot saves the weight parameters of the network in the best state. By rolling back to this state, problems such as overfitting during overtraining of the network can be avoided, ensuring the performance of the network.

[0101] Step S260: Input the optimal weight snapshot into the dynamic quantization encoder to generate network weight parameters for deployment encoded as 8-bit integers.

[0102] Step S270: Input the network weight parameters for deployment into the online incremental learning module, and update the parameters of the fully connected layer through clinical feedback data to adapt to the imaging characteristic differences of ultrasound probes.

[0103] The dynamic quantization encoder quantizes the weight parameters in the optimal weight snapshot. In the original network weight parameters, they may be represented by a high-precision data type (such as 32-bit floating-point numbers). The dynamic quantization encoder converts these weight parameters into 8-bit integer codes. For example, for a weight parameter representing the classification features of liver lesions, the original value is 0.12345 (represented by 32-bit floating-point numbers), and after being processed by the dynamic quantization encoder, it is converted into a value with an 8-bit integer code (the specific conversion process is based on the quantization algorithm). The network weight parameters with the generated 8-bit integer codes for deployment greatly reduce the storage requirements and computational volume of the network while maintaining a certain degree of accuracy, making the network more suitable for deployment in actual medical devices (such as ultrasound diagnostic devices). The network weight parameters for deployment are input into the online incremental learning module, and the parameters of the fully connected layer are updated through clinical feedback data to adapt to the imaging characteristic differences of ultrasound probes. In actual clinical applications, different ultrasound probes may have different imaging characteristics, such as differences in imaging resolution, imaging clarity of different tissues, etc. The clinical feedback data includes liver ultrasound images obtained according to the imaging of different ultrasound probes during actual use and the corresponding diagnostic results. Through these data, the parameters of the fully connected layer are updated. For example, if the display of the liver edge in the liver ultrasound image obtained by a certain ultrasound probe is not clear enough, resulting in some deviations during the classification process, then the online incremental learning module will adjust the weight parameters in the fully connected layer according to these clinical feedback data, enabling the network to better adapt to the imaging characteristic differences of different ultrasound probes and improving the accuracy of diagnosis.

[0104] In a possible implementation manner, the method further includes:

[0105] Step S310, comparing the output confidence entropy value of the cascaded deep classification network with a preset threshold, triggering an uncertainty warning event, inputting the uncertainty warning event into the generative adversarial network reconstruction module, calculating the reconstruction error distribution of the current ultrasound image, and performing conflict detection between the reconstruction error distribution and the classification probability distribution to locate the artifact interference area and trigger a standby classification model switching instruction.

[0106] The confidence entropy value reflects the uncertainty degree of the cascaded deep classification network for the classification result. For example, in the classification of liver ultrasound images, if the probability that the network outputs a certain area of the liver as normal is 0.8 and the probability of being a lesion is 0.2, according to the entropy calculation formula (-Σp(i)*log(p(i)), where p(i) is the probability of each category), the confidence entropy value can be calculated. Assuming the preset threshold is 0.5, when the calculated confidence entropy value is greater than the preset threshold, an uncertainty warning event is triggered. This means that the network is not very certain about the classification result of the current liver ultrasound image, and there may be some interference factors or limitations of the network itself.

[0107] For liver ultrasound images, assuming the image size is 500×600 pixels, each pixel has its corresponding feature value. The reconstruction error distribution is obtained by using a generative adversarial network to attempt to reconstruct the original image and calculating the difference between the reconstructed image and the original image. For example, within a 10×10 pixel sub-region in the liver area, the average pixel value of the liver tissue in the original image is 100 (assuming this value represents a certain feature intensity of the liver tissue), while the average pixel value of the corresponding region in the reconstructed image is 110. Then there is a certain reconstruction error in this region. The conflict detection is performed between this reconstruction error distribution and the classification probability distribution. If in a region with a high classification probability (such as 0.8 for normal), the reconstruction error is very large, this may imply the presence of artifact interference. Through this conflict detection, the artifact interference region can be located. For example, a 20×20 pixel region in the liver image is located as the artifact interference region, and its coordinate range is ((x1, y1)-(x2, y2)), and then the standby classification model switching instruction is triggered.

[0108] Step S320, input the coordinates of the artifact interference region and the model switching instruction into the self-healing log construction module, record the input data features and the output distribution deviation degree, generate a self-healing log, input the self-healing log into the graph neural network trainer, construct a fault knowledge graph of abnormal event association relationships, and input the fault knowledge graph into the network architecture optimizer to dynamically adjust the number of attention heads and the convolution kernel size of the cascaded deep classification network.

[0109] For the input data features, the features of the liver ultrasound image near the artifact interference region will be recorded, such as hemodynamic feature parameters, tissue elasticity parameters, etc. Assuming that near the artifact interference region, the measured value of the blood flow velocity is 40 cm / s (this value is extracted from the original ultrasound image data), and the tissue elasticity parameter is a specific value. The output distribution deviation degree refers to the degree of difference between the classification result under the condition of artifact interference and the classification result under the normal non-interference condition. For example, under normal circumstances, the probability that this region of the liver is classified as normal is 0.9, and when there is artifact interference, the probability of being classified as possibly diseased is 0.6. This change in probability is the output distribution deviation degree. By recording this information, a self-healing log is generated.

[0110] The graph neural network trainer analyzes the data in the self-healing log to find the correlation relationships between different abnormal events (such as artifact interferences at different positions, output distribution deviation degrees at different levels, etc.). For example, it is found that when the artifact interference area is located at the edge of the liver and the blood flow velocity measurement value is between 30 - 50 cm / s, it often leads to the classification result shifting towards the diseased direction, and this correlation relationship is constructed into the fault knowledge graph. The fault knowledge graph represents these correlation relationships in a graphical way. The nodes may represent different abnormal events or data features, and the edges represent the correlation strength between them.

[0111] For the cascaded deep classification network, the number of attention heads and the convolutional kernel size have an important impact on the performance of the network. If the fault knowledge graph shows that under certain specific liver ultrasound image features (such as specific hemodynamic features and artifact interference patterns), the classification effect of the network is not good, the network architecture optimizer will adjust the number of attention heads according to this information. Suppose the original number of attention heads is 8. According to the analysis of the fault knowledge graph, it is found that reducing the number of attention heads to 6 may improve the performance of the network in this case. For the convolutional kernel size, for example, the original convolutional kernel size is 3×3 pixels, and it may be adjusted to 5×5 pixels to better capture the features in the liver ultrasound image, especially the features near the artifact interference area.

[0112] Step S330: Input the optimized network architecture parameters into the full-scale self-check module. After repairing the numerical overflow nodes in the weight matrix and resetting the internal state parameters of the optimizer, synchronize the repaired network architecture parameters and the self-healing rule library to the distributed edge nodes to maintain the consistency of the abnormal handling strategy among multiple devices.

[0113] Finally, in the weight matrix, there may be numerical overflow nodes due to previous adjustments or numerical overflow problems during the calculation process. For example, a certain weight value should originally be between -1 and 1, but due to calculation errors or data type limitations, it becomes a very large value (such as 1000), and this is a numerical overflow node. The full-scale self-check module will repair these numerical overflow nodes and restore them to a reasonable numerical range. At the same time, reset the internal state parameters of the optimizer to ensure that the optimizer can perform subsequent training or adjustments in the correct state. Then, synchronize the repaired network architecture parameters and the self-healing rule library to the distributed edge nodes. In the diagnostic system of multiple devices, the distributed edge nodes may be distributed in different medical sites (such as different hospital departments or different regional medical centers). By synchronizing the repaired network architecture parameters and the self-healing rule library, each distributed edge node can maintain a consistent abnormal handling strategy, ensuring that the diagnostic results and handling methods among different devices are consistent when processing liver ultrasound images, and improving the reliability and accuracy of the entire diagnostic system.

[0114] In a possible implementation, after the step S150, the method further includes:

[0115] Step S410: Receive an interaction instruction triggered by a user in the interactive 3D visualization interface through a time axis control or a real-time ultrasound playback function, and parse the parameter adjustment request and the re-annotation coordinates of the lesion area in the interaction instruction.

[0116] For example, when a doctor views the visualization interface of a liver patient, using the time axis control to go back to an ultrasound image at a previous time point, or re-viewing a certain segment during the scanning process through the real-time ultrasound playback function, this will generate an interaction instruction. Thus, the parameter adjustment request and the re-annotation coordinates of the lesion area in this interaction instruction can be parsed. Suppose the doctor finds that there may be some inaccuracies in the previous diagnosis during the viewing process and wants to adjust the decision boundary of the classification probability distribution. The parameter adjustment request in the interaction instruction contains information about the confidence threshold offset. For example, the original confidence threshold for determining a certain area of the liver as an abnormal area is 0.7, and the doctor adjusts this threshold to 0.6 through the interaction instruction. This is a case where the confidence threshold offset is -0.1. At the same time, if the doctor finds that the previously marked coordinates of the lesion area are inaccurate and re-marks the lesion area, the new coordinate information is also included in the interaction instruction.

[0117] Step S420: Modify the decision boundary of the classification probability distribution according to the confidence threshold offset in the parameter adjustment request, generate a modified abnormal area overlay, and perform a topological consistency check on the re-annotation coordinates of the lesion area and the spatial topology relationship map, and update the blood vessel branch connection path in the 3D volume rendering model.

[0118] In liver ultrasound images, the classification probability distribution originally determines the probability of each region being an abnormal region. Taking a region in the right lobe of the liver as an example, the previously calculated probability of it being an abnormal region was 0.65, and it was determined to be a normal region when the confidence threshold was 0.7. However, when the confidence threshold was adjusted to 0.6, this region was included in the abnormal regions. The entire liver image is re-determined in this way to generate a corrected abnormal region overlay. In this overlay, the color coding may change according to the new determination results. For example, a region that was originally green (indicating normal) may become yellow (indicating suspected abnormality). At the same time, the topological consistency check is performed between the re-labeled coordinates of the lesion region and the spatial topological relationship atlas. The spatial topological relationship atlas contains the spatial relationships between blood vessels, tissues, and lesions in the liver. For the re-labeled coordinates of the lesion region, it is necessary to check whether its topological relationship with other structures in the atlas is reasonable. For example, whether the relative positional relationship between the lesion region and the surrounding blood vessels conforms to the physiological structure. If the re-labeled lesion region is close to a certain blood vessel, but in the previous spatial topological relationship atlas, the distance between this region and this blood vessel was far, further inspection and adjustment are required. Update the blood vessel branch connection path in the three-dimensional volume rendering model according to the verification results. If the position of the lesion region changes, it may affect the connection relationship with the blood vessel branches. For example, if it was originally considered that the lesion had no association with a certain small blood vessel branch, but it is found that there is an association after re-labeling and verification, the blood vessel branch connection path needs to be updated in the three-dimensional volume rendering model to accurately reflect the internal structural relationship of the liver.

[0119] Step S430, re-register the corrected abnormal region overlay with the updated three-dimensional volume rendering model to generate a secondary verification visualization framework, and extract the change gradient of the hemodynamic characteristic parameters in the secondary verification visualization framework.

[0120] During this process, it is necessary to ensure that each pixel point in the corrected abnormal area overlay accurately matches the corresponding position in the updated three-dimensional volume rendering model. For example, the area marked as yellow (suspected abnormality) in the corrected abnormal area overlay should also be accurately located at the corresponding position within the liver in the three-dimensional volume rendering model. Through this re-registration, the generated secondary verification visualization framework integrates the corrected two-dimensional probability distribution information and the updated three-dimensional anatomical structure information. Subsequently, the change gradient of the hemodynamic characteristic parameters in the secondary verification visualization framework is extracted. In liver images, the hemodynamic characteristic parameters include blood flow velocity, blood flow volume, etc. Suppose that in the previous visualization framework, the blood flow velocity of a major blood vessel in the liver was 50 cm / s, and in the secondary verification visualization framework, due to the re-annotation of the lesion area or other factors, the blood flow velocity becomes 45 cm / s. Then the change gradient of the blood flow velocity is -5 cm / s. Similar calculations are also performed for other hemodynamic characteristic parameters such as blood flow volume to obtain the change gradient of the hemodynamic characteristic parameters in the entire secondary verification visualization framework.

[0121] Step S440: Based on the change gradient of the hemodynamic characteristic parameters, perform incremental retrieval in the medical image database, screen out the historical case subset that matches the corrected classification probability distribution, and recalculate the risk level weight coefficient in the case similarity comparison result.

[0122] The medical image database stores a large amount of historical case images and their related characteristics of liver diseases. Taking the change gradient of the blood flow velocity as -5 cm / s as an example, search for those historical cases with similar changes in hemodynamic characteristic parameters in the database. Suppose there are 1000 historical cases in the database. By analyzing the hemodynamic characteristic parameters of each case, screen out the cases with a change gradient of hemodynamic characteristic parameters similar to that in the current secondary verification visualization framework. For example, screen out those historical cases with a blood flow velocity change within the range of -4 to -6 cm / s to form a historical case subset that matches the corrected classification probability distribution. Then recalculate the risk level weight coefficient in the case similarity comparison result. In the previous case similarity comparison, the risk level weight coefficient may have been determined based on factors such as liver size and lesion location. Now, after adding the new factor of the change gradient of hemodynamic characteristic parameters, re-evaluate the risk level weight coefficient. For example, for a historical case, the risk level weight coefficient originally calculated based on liver size and lesion location was 0.6. After considering the change gradient of hemodynamic characteristic parameters, it may be adjusted to 0.7 or 0.5, and the specific value depends on the impact of the new factor on the overall risk assessment.

[0123] Step S450, adjust the priority ranking of the treatment suggestions in the logical association table according to the recalculated risk level weight coefficients, generate a dynamically optimized treatment suggestion flowchart, and superimpose the dynamically optimized treatment suggestion flowchart on the secondary verification visualization framework to generate a final diagnosis report confirmed by the physician.

[0124] The logical association table originally determined different treatment suggestions and their priorities based on factors such as risk levels. For example, for liver lesions with a medium risk level, the previous treatment suggestion might be to first conduct drug treatment, then observe for a period of time, and if the condition does not improve, consider interventional treatment. This was the priority determined based on the previous risk level weight coefficients. When the risk level weight coefficients change, such as becoming a higher risk level, the priority of the treatment suggestion might be adjusted to directly perform interventional treatment first, and then assist with drug treatment. In this way, adjust the priority ranking of the treatment suggestions in the logical association table according to the recalculated risk level weight coefficients to generate a dynamically optimized treatment suggestion flowchart. Superimpose this dynamically optimized treatment suggestion flowchart on the secondary verification visualization framework to generate a final diagnosis report confirmed by the physician. In this final diagnosis report, the doctor can simultaneously see the optimized treatment suggestions and information such as liver images, structural relationships, and relevant characteristic parameters in the secondary verification visualization framework, facilitating the doctor's final diagnosis confirmation.

[0125] Step S460, input the correction parameters and relabeled coordinates in the final diagnosis report confirmed by the physician into the online incremental learning module, update the weights of the fully connected layer of the cascaded deep classification network, and synchronize the updated weights of the fully connected layer to the backup classification model of the distributed edge node to complete the cross-device diagnostic strategy consistency calibration.

[0126] For example, correction parameters such as the confidence threshold offset, the change gradient of hemodynamic characteristic parameters, and the relabeled lesion area coordinates in the final diagnosis report are input into the online incremental learning module. The online incremental learning module updates the weights of the fully connected layer of the cascaded deep classification network based on this new information. Suppose in the fully connected layer, the original value of a certain weight is 0.5, and during the update process, it might be adjusted to 0.6 or 0.4 according to the influence of the correction parameters. The updated weights of the fully connected layer are synchronized to the backup classification model of the distributed edge node. In a multi-device diagnostic system, the distributed edge nodes are distributed in different locations, such as different hospital departments or different regional medical centers. By synchronizing the updated weights, ensure that each device adopts the same diagnostic strategy when performing liver ultrasound image diagnosis, complete the cross-device diagnostic strategy consistency calibration, and improve the accuracy and reliability of the entire diagnostic system.

[0127] In a possible implementation, after the step of completing the cross-device diagnostic strategy consistency calibration, the method further includes:

[0128] Step S510, periodically collect multi-phase ultrasound reexamination data of the target object during the treatment stage, perform noise suppression and motion artifact compensation processing on each phase of the ultrasound reexamination data for the real-time ultrasound scan signal sequence, and generate a standardized follow-up ultrasound image sequence.

[0129] For example, it is set to perform a liver ultrasound reexamination on the patient every two weeks, and the ultrasound data obtained each time is one phase of the multi-phase ultrasound reexamination data. For each phase of the ultrasound reexamination data, perform the noise suppression and motion artifact compensation processing for the real-time ultrasound scan signal sequence before to generate a standardized follow-up ultrasound image sequence.

[0130] When performing noise suppression and motion artifact compensation processing, for each frame of the ultrasound scan signal in one phase of the ultrasound reexamination data, first perform wavelet transform decomposition. Assume that the size of the liver image corresponding to the ultrasound scan signal is 500×600 pixels (this pixel size is determined according to the imaging resolution of the ultrasound device). After wavelet transform decomposition, the high-frequency noise component and the low-frequency tissue signal component are separated. The high-frequency noise component shows some signal fluctuations in the high-frequency band, with relatively low energy and randomness, which may be caused by the electronic components of the ultrasound device, small interferences in the surrounding environment, etc. The low-frequency tissue signal component contains the general contour information of the main structures such as the liver tissue and blood vessels.

[0131] Then, use the adaptive threshold algorithm to perform layer-by-layer filtering on the high-frequency noise component. For example, for the first layer of the high-frequency noise component, analyze statistical parameters such as the standard deviation and mean of the signal in this layer. If the standard deviation is 5 (assume the unit), then according to a certain algorithm rule (such as multiplying the standard deviation by a coefficient, assume the coefficient is 3), the threshold for this layer is obtained as 15. The signal part higher than this threshold is regarded as a possible useful signal and retained, and the part lower than the threshold is determined as noise and filtered out. During the entire layer-by-layer filtering process, this adaptive threshold adjustment can effectively remove noise while retaining the intermediate-frequency signal component related to the movement of the blood vessel wall. For the blood vessels in the liver, the signal frequency generated by their blood flow is in the intermediate-frequency range, for example, between 1 - 5 MHz (assume), and these intermediate-frequency signals are of great significance for subsequent analysis of the movement state of the blood vessel wall, blood flow velocity, etc.

[0132] Then, perform inverse wavelet transform reconstruction on the filtered ultrasonic scan signal to generate a preliminarily denoised ultrasonic image. However, due to possible slight breathing or body movement of the patient during the reexamination scan, motion artifacts exist in the image. Calculate the tissue displacement vector field between adjacent frames of ultrasonic scan signals based on the optical flow method. For liver images, it is to calculate the movement direction and distance of liver tissue between adjacent frames. Assume that the time interval between two adjacent frames of images is 0.1 second (a typical time interval determined by the scan frame rate of the ultrasonic device). The optical flow method calculates the displacement vector of liver tissue within this 0.1 second. For example, the position of a specific point (assuming the coordinates are (x1, y1)) on the liver in the first frame of the image moves to (x2, y2) in the second frame of the image, then the calculated displacement vector is ((x2 - x1), (y2 - y1)). Perform motion trajectory compensation on the preliminarily denoised ultrasonic image according to this tissue displacement vector field. For example, a suspected lesion area in the liver has an obvious displacement in the image due to the patient's breathing movement before compensation. After motion trajectory compensation, the position of this area in the image can more accurately reflect its true position in the liver.

[0133] After that, perform gray normalization operation on the motion-compensated ultrasonic image to map the pixel intensity to a preset standard dynamic range. Assume that the preset standard dynamic range is 0 - 255 (a common pixel intensity range for images). In the motion-compensated liver ultrasonic image, first, statistically analyze the gray value range of all pixels in the image. For example, the gray value range of liver tissue may be 50 - 180, and the gray value range of blood vessels may be 30 - 100 (these are just assumed approximate ranges here). Then, through linear mapping, map these original gray values to the standard dynamic range of 0 - 255. The purpose of this is to make images of different periods and images of different patients comparable in terms of gray level, facilitating subsequent analysis and processing. After the gray normalization operation, a normalized ultrasonic image is obtained.

[0134] Finally, use the anisotropic diffusion algorithm to perform edge-preserving smoothing on the normalized ultrasonic image, and perform detail enhancement on the smoothed ultrasonic image through a bilateral filter. Then, perform region matching on the detail-enhanced ultrasonic image according to a preset anatomical structure template, and crop out a standardized follow-up ultrasonic image sequence containing the boundary of the complete liver organ.

[0135] Step S520: Input the standardized follow-up ultrasonic image sequence into the updated cascaded deep classification network, and output the follow-up classification probability distribution and spatial topology evolution map of the abnormal area in each follow-up ultrasonic image.

[0136] In a cascaded deep classification network, standardized follow-up ultrasound images for each period are analyzed. For example, for the follow-up ultrasound image of the liver in a certain period, the probability that a certain area of the liver is an abnormal area output by the network is 0.3 (this probability value is calculated based on the network's analysis of the image features), and this is a data point in the follow-up classification probability distribution. At the same time, the spatial topology evolution map can accurately depict the changes in the spatial relationship between the abnormal area and surrounding normal tissues, blood vessels and other structures during the follow-up process. For example, originally, the lesion area was close to a certain blood vessel in the liver at the initial diagnosis. During the follow-up process, as the treatment progresses, the distance between the lesion area and the blood vessel may change, and this change in the spatial relationship is recorded in the spatial topology evolution map.

[0137] Step S530: Align the follow-up classification probability distribution and the initial classification probability distribution in time series, calculate the abnormal area shrinkage rate and the blood flow signal intensity attenuation coefficient, and generate a set of treatment response quantification indicators.

[0138] Suppose the area of the liver lesion area in the initial classification probability distribution is calculated to be 100 square pixels through image analysis (assuming that one pixel represents an actual length of 0.1 mm, so the area is 1 square centimeter). In the follow-up classification probability distribution of a certain period, the area of the lesion area is calculated to be 80 square pixels (i.e., 0.8 square centimeters). Then the abnormal area shrinkage rate is (100 - 80) / 100 = 0.2, that is, 20%. For the blood flow signal intensity attenuation coefficient, for example, the blood flow signal intensity of a certain blood vessel in the liver lesion area at the initial diagnosis is 80 (assuming the intensity unit). In a certain period during the follow-up process, the blood flow signal intensity of this blood vessel in the lesion area becomes 60. Then the blood flow signal intensity attenuation coefficient is (80 - 60) / 80 = 0.25, that is, 25%. In this way, the abnormal area shrinkage rate and the blood flow signal intensity attenuation coefficient for each period are calculated, and a set of treatment response quantification indicators is generated.

[0139] Step S540: Input the spatial topology evolution map into a three-dimensional difference rendering engine, extract the blood vessel branch morphology change vector and the lesion infiltration volume change matrix, and generate a three-dimensional treatment response dynamic model.

[0140] In a three-dimensional differential rendering engine, based on the information in the spatial topology evolution atlas, for the blood vessel branches in the liver, analyze their morphological changes during the follow-up process. For example, the diameter of a certain blood vessel branch is 3 pixels at the initial diagnosis (assuming one pixel represents an actual length of 0.1 mm), and the length is 20 pixels (i.e., 2 mm). During the follow-up process, the diameter becomes 2.5 pixels and the length becomes 18 pixels. By calculating these change amounts, a morphological change vector of the blood vessel branch is obtained. For the lesion infiltration volume change matrix, assume that the lesion infiltration volume is calculated to be 100 cubic pixels (i.e., 1 cubic centimeter) through three-dimensional analysis of the lesion area at the initial diagnosis. During the follow-up process, according to the spatial topology evolution atlas of different stages, different lesion infiltration volumes are calculated. For example, in a certain stage, it is 80 cubic pixels (i.e., 0.8 cubic centimeter). These data are used to form the lesion infiltration volume change matrix. Through these data, a three-dimensional treatment response dynamic model is generated, and this three-dimensional treatment response dynamic model can visually display the three-dimensional structural changes of the liver during the treatment process.

[0141] Step S550: Integrate the set of treatment response quantification indicators and the three-dimensional treatment response dynamic model to generate a multi-stage treatment effect comparison report, and perform a prognosis pattern matching between the multi-stage treatment effect comparison report and the historical treatment cases in the medical image database.

[0142] The multi-stage treatment effect comparison report combines quantified indicators (such as the shrinkage rate of the abnormal area and the attenuation coefficient of the blood flow signal intensity) and a three-dimensional structural change model (the three-dimensional treatment response dynamic model). Then, a prognosis pattern matching is performed between the multi-stage treatment effect comparison report and the historical treatment cases in the medical image database. The medical image database stores a large number of historical treatment cases of liver diseases and their related treatment effect data. By matching the indicators and models in the multi-stage treatment effect comparison report with the historical treatment cases, the historical treatment case most similar to the current patient's treatment situation is found.

[0143] Step S560: Screen the best historical treatment case according to the prognosis pattern matching result, extract the corresponding treatment plan flowchart and key node monitoring parameters, and generate a personalized treatment monitoring dashboard.

[0144] For example, through prognosis pattern matching, it is found that a certain historical treatment case is the most similar to the current patient's treatment situation. The treatment plan flowchart of this historical treatment case includes information such as the types, dosages, treatment cycles of drug treatments, and the examination items required at different stages during the treatment process. The key node monitoring parameters may include the expected change ranges of liver function indicators (such as alanine aminotransferase, aspartate aminotransferase, etc.) and the expected changes in the size of the liver lesion area at a certain stage of the treatment. These information are extracted to generate a personalized treatment monitoring dashboard.

[0145] Step S570: Compare the monitoring parameter thresholds in the personalized treatment monitoring dashboard with the real-time ultrasound reexamination data stream in real time. When a deviation warning event is triggered, automatically adjust the drug dosage parameters and reexamination time interval in the treatment recommendation flowchart to generate an adaptive treatment plan update instruction.

[0146] For example, in the personalized treatment monitoring dashboard, it is set that the area of the liver lesion area should be reduced to a certain range at a certain stage of treatment. If it is found in the real-time ultrasound reexamination data stream that the area of the liver lesion area has not reached the expected reduction range, a deviation warning event is triggered. At this time, according to the pre-set rules, automatically adjust the drug dosage parameters in the treatment recommendation flowchart, such as increasing the dosage of a certain drug from 10 mg per day to 15 mg, and at the same time adjust the reexamination time interval, such as changing from the original reexamination every two weeks to once a week, to generate an adaptive treatment plan update instruction.

[0147] Step S580: Logically associate the adaptive treatment plan update instruction with the multi-phase treatment effect comparison report to generate a treatment closed-loop feedback file, and transmit the key decision parameters in the treatment closed-loop feedback file back to the online incremental learning module to optimize the subsequent classification network weight update direction.

[0148] The treatment closed-loop feedback file records the logical association relationship from the adaptive treatment plan update instruction to the multi-phase treatment effect comparison report, including the basis for adjusting the drug dosage and reexamination time interval, expected treatment effects and other information. Transmit the key decision parameters (such as drug dosage adjustment amount, reexamination time interval adjustment amount, etc.) in it back to the online incremental learning module, and the online incremental learning module optimizes the subsequent classification network weight update direction according to these parameters. For example, if it is found that the improvement of the liver lesion area is better after increasing the drug dosage, then the online incremental learning module will adjust the weights related to drug dosage and treatment effect in the cascaded deep classification network so that in subsequent diagnoses and treatment recommendations, this relationship can be more accurately reflected, improving the accuracy and adaptability of the entire diagnosis and treatment system.

[0149] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an artificial intelligence-based ultrasound image data classification system 100 that can implement the ideas of the present application provided by some embodiments of the present application. For example, the processor 120 can be used on the artificial intelligence-based ultrasound image data classification system 100 and is used to execute the functions in the present application.

[0150] The artificial intelligence-based ultrasonic image data classification system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the artificial intelligence-based ultrasonic image data classification method of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0151] For example, the artificial intelligence-based ultrasonic image data classification system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the artificial intelligence-based ultrasonic image data classification system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The artificial intelligence-based ultrasonic image data classification system 100 also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0152] For ease of illustration, only one processor is described in the artificial intelligence-based ultrasonic image data classification system 100. However, it should be noted that the artificial intelligence-based ultrasonic image data classification system 100 in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application can also be jointly executed or separately executed by multiple processors. For example, if the processor of the artificial intelligence-based ultrasonic image data classification system 100 executes steps A and B, it should be understood that steps A and B can also be jointly executed by two different processors or separately executed in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0153] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above artificial intelligence-based ultrasonic image data classification method is implemented.

[0154] It should be noted that, in order to simplify the expression of the disclosure of the present invention and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An artificial intelligence-based method for classifying ultrasonic image data, characterized in that, The method includes: Obtaining a real-time ultrasound scan signal sequence of a target object, wherein the real-time ultrasound scan signal sequence includes multiple frames of dynamic ultrasound images, and the temporal variation characteristics of tissue elasticity parameters and hemodynamic parameters are embedded in each frame of the dynamic ultrasound image; Performing noise suppression and motion artifact compensation processing on the real-time ultrasound scan signal sequence to generate a denoised standardized ultrasound image sequence, and marking the anatomical boundary coordinates of the region of interest in the standardized ultrasound image sequence; Performing multi-scale anatomical structure decomposition on each frame of the standardized ultrasound image based on the anatomical boundary coordinates to generate a set of local feature maps corresponding to different tissue levels, and each local feature map in the set of local feature maps includes a three-dimensional matrix of edge sharpness index, texture consistency index, and blood flow signal intensity distribution; Inputting the set of local feature maps into a cascaded deep classification network, and through the spatio-temporal feature alignment module and multi-granularity attention allocation module nested in the cascaded deep classification network, performing cross-frame feature fusion and dynamic weight adjustment on the set of local feature maps, and outputting the classification probability distribution and spatial topology relationship map of the abnormal region in each frame of the standardized ultrasound image; Generating a multi-modal diagnostic report according to the classification probability distribution and the spatial topology relationship map, and comparing the similarity of the multi-modal diagnostic report with the historical case features in the medical image database to generate an interactive three-dimensional visualization interface including a risk level label and treatment suggestions.

2. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein The performing noise suppression and motion artifact compensation processing on the real-time ultrasound scan signal sequence to generate a denoised standardized ultrasound image sequence includes: Performing wavelet transform decomposition on each frame of the ultrasound scan signal in the real-time ultrasound scan signal sequence to separate the high-frequency noise component and the low-frequency tissue signal component; Using an adaptive threshold algorithm to perform layer-by-layer filtering on the high-frequency noise component to obtain a filtered ultrasound scan signal, and retaining the intermediate-frequency signal component related to the movement of the blood vessel wall; Performing inverse wavelet transform reconstruction on the filtered ultrasound scan signal to generate a preliminarily denoised ultrasound image; Calculating the tissue displacement vector field between adjacent frames of ultrasound scan signals based on the optical flow method, compensating the movement trajectory of the preliminarily denoised ultrasound image according to the tissue displacement vector field, and performing gray normalization operation on the motion-compensated ultrasound image to map the pixel intensity to a preset standard dynamic range to obtain a normalized ultrasound image; Performing edge-preserving smoothing processing on the normalized ultrasound image using an anisotropic diffusion algorithm, enhancing the details of the smoothed ultrasound image through a bilateral filter, and then performing region matching on the detail-enhanced ultrasound image according to a preset anatomical structure template to crop out a standardized ultrasound image sequence including the complete organ boundary.

3. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein, The performing multi-scale anatomical structure decomposition on each frame of the standardized ultrasound image based on the anatomical boundary coordinates to generate a set of local feature maps corresponding to different tissue levels includes: In each frame of the standardized ultrasound image, intercepting a rectangular region of interest including the target organ according to the anatomical boundary coordinates; Perform Gaussian pyramid decomposition on the rectangular region of interest to generate a multi-scale image group including the original resolution, half resolution, and quarter resolution images; Apply a multi-directional Gabor filter bank to each scale image in the multi-scale image group to extract the frequency domain response feature maps that match the blood vessel branch directions; Input the frequency domain response feature maps of different scales into a three-dimensional convolutional neural network, aggregate the edge gradient information at different resolutions through a cross-scale feature fusion layer, and perform non-maximum suppression processing on the aggregated feature maps to eliminate redundant edge responses and retain the maximum intensity edge points. Then, based on the retained edge points, construct a spatial Delaunay triangular mesh, and calculate the curvature distribution and area ratio index of each triangular patch of the triangular mesh; Perform regional division on the triangular mesh according to the curvature distribution and area ratio index to generate a sub-grid set corresponding to different tissue types, and statistically calculate the spatio-temporal variation variance of the blood flow signal intensity within each sub-grid of the sub-grid set to generate a local feature map set including edge geometric features and hemodynamic features.

4. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein Input the local feature map set into a cascaded deep classification network. Through the spatio-temporal feature alignment module and multi-granularity attention allocation module nested in the cascaded deep classification network, perform cross-frame feature fusion and dynamic weight adjustment on the local feature map set, and output the classification probability distribution and spatial topology relationship map of the abnormal regions in each frame of the standardized ultrasound image, including: Input the local feature map set into the long short-term memory network of the spatio-temporal feature alignment module for temporal state transfer to generate a time dimension feature vector including the morphological evolution law of the lesion between consecutive frames, and input the time dimension feature vector into a deformable convolutional layer for spatial deformation modeling to output a spatial deformation feature matrix that matches the non-linear deformation of the tissue structure; Concatenate the time dimension feature vector and the spatial deformation feature matrix to generate a spatio-temporal joint feature matrix, and input the spatio-temporal joint feature matrix into the multi-head self-attention mechanism of the multi-granularity attention allocation module to calculate the global correlation weights between different feature channels and output a global attention weight matrix; Input the global attention weight matrix into a gated linear unit for feature channel screening to retain the feature channels relevant to the current classification task, generate a gated screening feature vector, and input the gated screening feature vector into a local window attention calculation layer to enhance the spatial correlation of adjacent feature nodes and generate a local window attention weight matrix; Fusion the global attention weight matrix and the local window attention weight matrix with weights to generate a multi-granularity attention allocation map, and superimpose the multi-granularity attention allocation map on the spatio-temporal joint feature matrix through a residual connection to generate an optimized classification feature vector; Input the optimized classification feature vector into a fully connected classification layer to output the classification probability distribution and spatial topology relationship map of the abnormal regions in each frame of the standardized ultrasound image.

5. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein, The method generates a multimodal diagnosis report based on the classification probability distribution and the spatial topological relationship map, and compares the multimodal diagnosis report with the historical case characteristics in the medical imaging database for similarity, and generates an interactive three-dimensional visualization interface including risk level labels and treatment recommendations, including: Mapping the classification probability distribution to the spatial coordinates of a standardized ultrasound image sequence to generate an abnormal region overlay with semi-transparent color coding; Inputting the spatial topological relationship map into a three-dimensional volume rendering engine to construct a three-dimensional volume rendering model with annotated blood vessel branches and lesion infiltration ranges; Performing spatial registration of the abnormal area cover layer with the three-dimensional volume rendering model to generate a visualization basic framework integrating the two-dimensional probability distribution and the three-dimensional anatomical structure; Extracting the hemodynamic characteristic parameters in the classification probability distribution, retrieving historical case characteristics with similar hemodynamic characteristic parameters in the medical imaging database, and generating case similarity comparison results; The case similarity comparison result is associated with the visualization basic framework to generate a side comparison view and a key difference parameter annotation layer, and the preset treatment plan flow chart is matched based on the key difference parameter annotation layer to generate a logical association table of risk level labels and treatment recommendations. The logical association table is then embedded in the visualization basic framework to generate an interactive three-dimensional visualization interface including an interactive timeline control, a real-time ultrasound playback function and a voice broadcast engine.

6. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein The method further comprises: The training sample library with multi-dimensional labels is input into the data augmentation module, and random elastic deformation, local pixel occlusion and multi-angle affine transformation are performed to generate an expanded adversarial training sample set; Inputting the expanded adversarial training sample set into the focal loss function of the first training stage to optimize the inter-class balance weight parameters of the cascaded deep classification network; The optimized inter-class balance weight parameters are input into the adversarial training strategy of the second training phase, and an adversarial perturbation sample set is generated through gradient sign attack. Inputting the adversarial perturbation sample set into the knowledge distillation framework of the third training stage, and utilizing the soft label transfer mechanism of the teacher network to compress the model complexity parameters of the student network; The compressed model complexity parameters are input into the validation set monitoring module, and the early stopping mechanism is triggered according to the classification accuracy and generalization performance, and the model is rolled back to the optimal weight snapshot. Inputting the optimal weight snapshot into a dynamic quantization encoder to generate an 8-bit integer-encoded network weight parameter for deployment; The network weight parameters for the deployment are input into an online incremental learning module, and the fully connected layer parameters are updated through clinical feedback data to adapt to the differences in ultrasound probe imaging characteristics.

7. The method for classifying ultrasonic image data based on artificial intelligence according to claim 6, characterized in that, The method further comprises: Compare the output confidence entropy value of the cascaded deep classification network with a preset threshold value to trigger an uncertainty warning event, and input the uncertainty warning event into a generative adversarial network reconstruction module to calculate the reconstruction error distribution of the current ultrasound image, and perform conflict detection between the reconstruction error distribution and the classification probability distribution, locate the artifact interference area and trigger a backup classification model switching instruction; Input the coordinates of the artifact interference region and the model switching instruction into the self-healing log construction module, record the characteristics of the input data and the deviation degree of the output distribution, generate a self-healing log, input the self-healing log into the graph neural network trainer to construct a fault knowledge graph of abnormal event association relationships, and input the fault knowledge graph into the network architecture optimizer to dynamically adjust the number of attention heads and the convolution kernel size of the cascaded deep classification network; Input the optimized network architecture parameters into the full-scale self-check module. After repairing the numerical overflow nodes in the weight matrix and resetting the internal state parameters of the optimizer, synchronize the repaired network architecture parameters and the self-healing rule library to the distributed edge nodes to maintain the consistency of the abnormal handling strategies among multiple devices.

8. The method for classifying ultrasonic image data based on artificial intelligence according to claim 1, wherein After generating the interactive three-dimensional visualization interface including risk level labels and treatment suggestions, the method further includes: Receive the interaction instructions triggered by the user in the interactive three-dimensional visualization interface through the time axis control or the real-time ultrasound playback function, and parse the parameter adjustment request and the re-annotated coordinates of the lesion area in the interaction instructions; Modify the decision boundary of the classification probability distribution according to the confidence threshold offset in the parameter adjustment request, generate a modified abnormal area overlay, and perform a topological consistency check on the re-annotated coordinates of the lesion area and the spatial topology relationship graph to update the blood vessel branch connection path in the three-dimensional volume rendering model; Re-register the modified abnormal area overlay and the updated three-dimensional volume rendering model to generate a secondary verification visualization framework, and extract the change gradient of the hemodynamic characteristic parameters in the secondary verification visualization framework; Perform incremental retrieval in the medical image database based on the change gradient of the hemodynamic characteristic parameters, filter out the historical case subset that matches the modified classification probability distribution, and recalculate the risk level weight coefficient in the case similarity comparison result; Adjust the priority sorting of the treatment suggestions in the logical association table according to the recalculated risk level weight coefficient, generate a dynamically optimized treatment suggestion flow chart, and overlay the dynamically optimized treatment suggestion flow chart and the secondary verification visualization framework to generate the final diagnosis report confirmed by the physician; Input the correction parameters and re-annotated coordinates in the final diagnosis report confirmed by the physician into the online incremental learning module to update the weights of the fully connected layer of the cascaded deep classification network, and synchronize the updated weights of the fully connected layer to the backup classification model of the distributed edge node to complete the calibration of the cross-device diagnosis strategy consistency.

9. The method for classifying ultrasonic image data based on artificial intelligence according to claim 8, wherein After the step of completing the cross-device diagnosis strategy consistency calibration, the method further includes: Periodically collect the multi-phase ultrasound reexamination data of the target object during the treatment stage, perform noise suppression and motion artifact compensation processing on the real-time ultrasound scanning signal sequence of each phase of the ultrasound reexamination data to generate a standardized follow-up ultrasound image sequence; Input the standardized follow-up ultrasound image sequence into the updated cascaded deep classification network to output the follow-up classification probability distribution of the abnormal area and the spatial topology evolution graph in each follow-up ultrasound image; Perform time series alignment on the follow-up classification probability distribution and the initial classification probability distribution, calculate the shrinkage rate of the abnormal area and the attenuation coefficient of the blood flow signal intensity, and generate a set of treatment response quantification indicators; Input the spatial topological evolution atlas into a three-dimensional differential rendering engine, extract the change vector of the blood vessel branch morphology and the change matrix of the lesion infiltration volume, and generate a three-dimensional treatment response dynamic model; Fuse the set of treatment response quantification indicators and the three-dimensional treatment response dynamic model to generate a multi-phase treatment effect comparison report, and perform prognostic pattern matching on the multi-phase treatment effect comparison report and the historical treatment cases in the medical image database; Screen the best historical treatment cases according to the prognostic pattern matching results, extract the corresponding treatment plan flow chart and key node monitoring parameters, and generate a personalized treatment monitoring dashboard; Compare the monitoring parameter thresholds in the personalized treatment monitoring dashboard with the real-time ultrasound reexamination data stream in real time. When a deviation warning event is triggered, automatically adjust the drug dose parameters and reexamination time interval in the treatment recommendation flow chart to generate an adaptive treatment plan update instruction; Logically associate the adaptive treatment plan update instruction with the multi-phase treatment effect comparison report to generate a treatment closed-loop feedback file, and transmit the key decision parameters in the treatment closed-loop feedback file back to the online incremental learning module to optimize the subsequent classification network weight update direction.

10. An ultrasound image data classification system based on artificial intelligence, characterized in that, The artificial intelligence-based ultrasound image data classification system includes a processor and a memory. The memory is connected to the processor. The memory is used to store programs, instructions or codes. The processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based ultrasound image data classification method according to any one of claims 1-9 above.

Images