Deep learning based method for tfcc injury identification in wrist mri

Abstract

The present application relates to the field of medical imaging technology, and in particular to a TFCC injury identification method in a wrist joint MRI based on deep learning, wherein by analyzing the morphological shift and gradient change of feature points in the initial region of an MRI image in a continuous initial region image, the variation coefficients of each feature point in the initial region are obtained; then, according to the variation coefficients, the extension performance of the bony structure of the initial region is analyzed to obtain the regional variation cutoff coefficient of the initial region; then, the initial region is subjected to convolution analysis, and the attention performance of the initial region in each convolution layer is analyzed in combination with the regional variation cutoff coefficient; the attention weight of the initial region in each convolution layer is adjusted based on the attention performance; and finally, the convolution neural network is trained through the adjusted attention weight, which can effectively improve the feature extraction capability of the restrictive performance of different injury types, and thus effectively improve the accuracy of TFCC injury identification and classification in a wrist joint MRI.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, specifically to a deep learning-based method for identifying TFCC lesions in wrist MRI. Background Technology

[0002] The triangular fibrocartilage complex (TFCC) is a key fibrocartilage-ligament complex located on the ulnar side of the wrist joint. It primarily stabilizes the distal radioulnar and carpal joints, while also transmitting axial loads and allowing for forearm rotation. The Palmer classification system classifies TFCC injuries into two main categories based on the injury mechanism. MRI shows discontinuity in the TFCC, with high signal intensity on various sequences, especially T2-weighted images (T2WI) and proton density-weighted fat-suppressed sequences (PD-FS). Figure 2 and Figure 3 (As shown), there is high signal effusion between the distal radiocarpal joint and the radiolunate joint. The first type is traumatic tear caused by acute injury, such as a fall where the hand hits the ground, with forearm pronation and wrist hyperextension. This is common in young people and athletes, the injury site is clear, and it can often be repaired surgically. The second type is degenerative injury caused by chronic wear and tear, often related to positive ulnar variations, and is more common in people over 40 years old. It manifests as progressive ulnar pain, and treatment focuses on decompression and debridement. Normal TFCC shows a uniform low signal, continuous structure, and no high signal areas. Because the treatment plans for the two types of TFCC injuries are different, it is necessary to identify the type of TFCC injury in clinical practice. Detecting the type of TFCC injury is a clinical challenge, as the two types of TFCC injuries appear similar in routine examinations and imaging studies, thus affecting the accuracy of identification. Summary of the Invention

[0003] To address the above technical problems, this invention provides a deep learning-based method for identifying TFCC lesions in wrist joint MRI, comprising: Obtain MRI images of the patient's wrist joint; For the MRI image, initial region identification is performed to obtain multiple consecutive initial region images. Then, the morphological shift and gradient change of feature points in the initial region in the consecutive initial region images are analyzed to obtain the variation coefficient of each feature point in the initial region. Based on the variation coefficient, the extension performance of the bony structure in the initial region is analyzed to obtain the regional variation cutoff coefficient of the initial region. Convolutional analysis is performed on the initial region to obtain the feature maps of each convolutional layer. Combined with the regional variation cutoff coefficient, a regional feature coefficient map of the initial region is constructed. Then, the significant performance of the initial region in each convolutional layer is analyzed to obtain the attention performance of the initial region in each convolutional layer. Based on the attention performance, the minimum difference between the initial region and other initial regions is analyzed, the attention weights of the initial region in each convolutional layer are adjusted, and then a convolutional neural network is trained to achieve TFCC damage recognition.

[0004] In some embodiments of the present invention, initial region identification is performed on the MRI image to obtain multiple consecutive initial region images, including: For the MRI image, threshold segmentation is performed to identify the initial regions corresponding to the lunate bone, triquetrum bone, styloid process of the ulna, and TFCC of the wrist joint, and multiple continuous initial region images are obtained, where multiple continuous initial region images refer to continuous slices of the MRI image on the Z-axis.

[0005] In some embodiments of the present invention, the morphological shift and gradient change of feature points in the initial region in continuous initial region images are analyzed to obtain the variation coefficients of each feature point in the initial region, including: Feature points are extracted from multiple consecutive initial region images, and the feature points in the multiple consecutive initial region images are matched. For the matched feature points, the grayscale differences of the feature points in the initial region are analyzed in all adjacent initial region images to obtain the gradient change value of each feature point in the initial region in all adjacent initial region images. For the matched feature points, the differences in coordinate points of the feature points in the initial region among all adjacent initial region images are analyzed to obtain the morphological offset value of each feature point in the initial region among all adjacent initial region images. Based on the gradient change value and the shape offset value, the variation coefficient of each feature point in each initial region is obtained.

[0006] In some embodiments of the present invention, the extension behavior of the bony structure in the initial region is analyzed based on the variation coefficient to obtain the regional variation cutoff coefficient of the initial region, including: Based on the variation coefficient and the coordinate data of the feature points, the extension direction of the bony structure of the initial region is obtained, and then curve fitting and curve segmentation are performed to obtain multiple sub-curves corresponding to each initial region. Based on the mean variation coefficient of each sub-curve, cluster analysis is performed on all the sub-curves corresponding to each initial region to obtain 3 clusters; The average variation coefficient of each cluster is obtained, and the differences in the average variation coefficients between clusters are analyzed to obtain the regional variation cutoff coefficient of each initial region.

[0007] In some embodiments of the present invention, the extension direction of the bony structure of the initial region is obtained based on the variation coefficient and the coordinate data of the feature points, and then curve fitting and curve segmentation are performed to obtain multiple sub-curves corresponding to the initial region, including: Based on the variation coefficient, PCA principal component analysis is performed to obtain the first principal component direction as the extension direction of the bony structure of the initial region; Using the extension direction as the horizontal axis and the direction perpendicular to the extension direction as the vertical axis, the x-values ​​of the projection coordinates of all feature points in the initial region on the horizontal axis are used as the horizontal axis coordinates, and the variation coefficients of all feature points in the initial region are used as the vertical axis coordinates. Curve fitting is performed on all feature points in the initial region to obtain the fitting curve for each initial region. The curve is segmented based on the slope of all data points on the fitted curve to obtain multiple sub-curves corresponding to each initial region.

[0008] In some embodiments of the present invention, convolutional analysis is performed on the initial region to obtain feature maps of each convolutional layer, and a regional feature coefficient map of the initial region is constructed by combining the regional variation cutoff coefficient, including: The initial region is analyzed by convolutional analysis using the U-Net fully convolutional neural network to obtain the feature map of each initial region in each convolutional layer; The regional variation cutoff coefficient of the initial region is regarded as the feature coefficient of all position pixels in the initial region, and a regional feature coefficient map of the same size as the feature map is constructed for each initial region.

[0009] In some embodiments of the present invention, the analysis of the significant performance of the initial region in each convolutional layer includes: By analyzing the differences between the region feature coefficient map of the initial region and the feature map of the initial region in each convolutional layer, the abnormal performance of the region feature coefficient map of each initial region in each convolutional layer is obtained. Based on the abnormal behavior, the differences in abnormal behavior of the initial region among all adjacent convolutional layers are analyzed, and the differences between the abnormal behavior of the region feature coefficient map of the initial region on each convolutional layer and the mean of abnormal behavior on all convolutional layers are analyzed to obtain the significant behavior of the initial region on each convolutional layer.

[0010] In some embodiments of the present invention, based on the attention performance, the minimum difference between the initial region and other initial regions is analyzed, and the attention weights of the initial region in each convolutional layer are adjusted, including: For each convolutional layer, the minimum absolute value of the difference between the attention performance of the initial region and the attention performance of all other initial regions is obtained, thus obtaining the minimum difference between the initial region and the other initial regions. Based on the minimum difference, and by introducing an adjustment coefficient and an inverse proportional exponential function, the attention weights of the initial region in each convolutional layer are adjusted to obtain the adjusted attention weights of the initial region in each convolutional layer.

[0011] In some embodiments of the present invention, convolutional neural network training is performed to achieve TFCC damage recognition, including: Based on the adjusted attention weights of all the initial regions in each convolutional layer, an attention weight matrix corresponding to the feature map size is obtained for each convolutional layer; The attention weight matrices of each convolutional layer are used to train the U-Net fully convolutional neural network to obtain the trained U-Net fully convolutional neural network. The MRI image is input into a trained U-Net fully convolutional neural network, and the output is a multi-channel probability map, with each channel corresponding to a type of damage; The multi-channel probability map is thresholded to generate a binary mask, which identifies the pixel position of each damaged area.

[0012] In some embodiments of the present invention, acquiring MRI images of a patient's wrist joint includes: Initial MRI images of the patient's wrist joint, including proton density-weighted fat-suppressed sequences and T2-weighted sequences, were acquired. For the initial MRI images, the proton density-weighted fat-suppressed sequence and the T2-weighted image sequence were rigidly registered to obtain the registered MRI images; For the registered MRI images, a three-dimensional local region is cropped with the distal ulnar head of the wrist joint as the center, focusing on the TFCC and the lunate, triquetrum, and ulnar styloid processes, and removing irrelevant tissues to obtain the MRI images of the patient's wrist joint.

[0013] This invention offers the following advantages: It identifies initial regions in MRI images, acquiring multiple consecutive initial region images. It then analyzes the morphological shifts and gradient changes of feature points within these initial regions, obtaining the variation coefficients of each feature point. These variation coefficients reflect the structural changes in the local area. Actual structural changes are influenced by the type of injury or alterations in bone structure, which are reflected in the correlation of the variation coefficients of the corresponding feature points. Based on these variation coefficients, the extension of the bony structure in the initial region is analyzed, yielding a regional variation cutoff coefficient that effectively reflects the characteristics of the corresponding injury type. Convolutional analysis is then performed on the initial region, and combined with the regional variation cutoff coefficient, the significant performance of the initial region in each convolutional layer is analyzed, revealing its attentional performance. Based on this attentional performance, the minimum difference between the initial region and other initial regions is analyzed, and the attention weights of the initial region in each convolutional layer are adjusted to enhance sensitivity to these key features. Finally, the adjusted attention weights are used to train a convolutional neural network, effectively improving the feature extraction capability of the restrictive representation of abnormal shapes under the actual structural relationships of different injury types. This significantly improves the accuracy of TFCC injury identification and classification in wrist MRI. Attached Figure Description

[0014] Figure 1 A flowchart of a deep learning-based method for identifying TFCC lesions in wrist MRI provided by this invention; Figure 2 A schematic diagram of a proton density-weighted lipid suppression sequence (PD-FS) high-brightness signal provided by the present invention; Figure 3 This is a schematic diagram of a T2-weighted image sequence (T2WI) highlight signal provided by the present invention. Detailed Implementation

[0015] Different types of TFCC injuries exhibit different MRI visual appearances, primarily characterized by high signal intensity located near different areas and accompanied by varying signs of complications. For example, in cases of joint effusion, a high-signal fluid shadow on T2WI is visible on the ulnar side of the wrist, suggesting synovitis or an inflammatory response; in cases of synovial thickening, the soft tissues surrounding the TFCC show nodular or diffuse high signal intensity. Both joint effusion and synovial thickening further influence the actual presentation of the injury. However, the abnormal signal intensity in different types of TFCC injuries is constrained by the inherent anatomical structure of the TFCC. The injury area cannot cross the tip of the ulnar styloid process or extend to the distal radius. For instance, in laceration-like injuries, the high signal intensity extends radially from the ulnar attachment point of the TFCC along the ligament fibers, existing only in the space between the bones. Therefore, by analyzing the restrictive regional manifestations of different types of TFCC injuries on MRI images, the structural feature attention in deep learning neural networks can be improved, effectively achieving accurate identification of different injury types.

[0016] The following description, in conjunction with the accompanying drawings, details a specific scheme for a deep learning-based TFCC lesion identification method in wrist MRI provided by the present invention.

[0017] Please see Figure 1 This illustrates the basic process of a deep learning-based TFCC lesion identification method in wrist MRI provided by an embodiment of the present invention.

[0018] like Figure 1 As shown, an embodiment of the present invention provides a deep learning-based method for identifying TFCC lesions in wrist joint MRI, specifically including: S100: Acquire MRI images of the patient's wrist joint.

[0019] Obtain MRI images of the patient's wrist joint, further including: First, initial MRI images of the patient's wrist joint were acquired, including proton density-weighted fat-suppressed sequences (PD-FS) and T2-weighted imaging sequences (T2WI). Specifically, a 3.0T MRI machine was used, with a slice thickness ≤2mm and slice interval ≤0.2mm to ensure sufficient spatial resolution to resolve micro-tears (<2mm). During MRI image acquisition, the patient's wrist joint was positioned in a neutral position with the forearm pronated to ensure the TFCC was under maximum tension to expose potential damage.

[0020] Then, for the initial MRI images, the proton density-weighted fat-suppressed sequence and the T2-weighted image sequence were rigidly registered to obtain the registered MRI images of the patient's wrist joint, ensuring that the same anatomical structure is spatially consistent in different sequences.

[0021] Finally, for the registered MRI images, a three-dimensional local region centered on the distal ulnar head of the wrist joint is cropped, focusing on the TFCC and the lunate, triquetrum, and ulnar styloid processes, and removing irrelevant tissues to obtain the cropped MRI image. Specifically, the pixel intensity of all registered MRI images is linearly mapped to the [0,1] interval to obtain a standardized registered MRI image, eliminating signal differences between devices; and for the standardized registered MRI image, a 64×64×32 voxel three-dimensional local region is cropped centered on the distal ulnar head of the wrist joint, focusing on the TFCC and the lunate, triquetrum, and ulnar styloid processes, and removing irrelevant tissues (such as forearm muscles and skin) to obtain the desired MRI image of the patient's wrist joint.

[0022] Since the identification of TFCC injury types relies more on the relative structural relationships and actual location of wrist joint tissues as shown in images—for example, ulnar tears are characterized by localized high signal intensity, with linear high signal intensity extending from the periphery of the TFCC to the articular surface on T2WI, indicating ligament avulsion from its attachment point to the ulna—these signs are not affected by the relative positional changes of the corresponding bones and soft tissues in the wrist joint. The presence of a bright, extended tear at that location indicates an injury. Therefore, different injury types are constrained by abnormal shapes under the actual structural relationships. Furthermore, due to the limitations imposed by the relationship loss in deep learning networks, further analysis and adjustment of the attention weights in the deep learning neural network are needed to achieve accurate identification of different injury types. In embodiments of this invention, steps S200 to S500 are specifically included.

[0023] S200: For MRI images, initial region identification is performed to obtain multiple consecutive initial region images. Then, the morphological shift and gradient change of feature points in the initial region in the consecutive initial region images are analyzed to obtain the variation coefficient of each feature point in the initial region.

[0024] The relative structure of the wrist joint shown on MRI images mainly refers to changes in the relative positions of the corresponding bones, ligaments, and soft tissues. In addition to the manifestations of different injury types, these changes can also be caused by mechanical imbalances in the overall structure of the wrist joint triggered or accompanied by TFCC injuries, which manifest as changes in bony relationships. These are indirect signs of injury and will actually change the relationship of the wrist joint structure on MRI images, that is, some bones and tissues will undergo morphological shifts.

[0025] However, the specific manifestations of different TFCC injury types are significantly regionally limited compared to changes in bony relationships. This is because the corresponding injury manifestations only exist in specific structures. For example, the high signal intensity of Palmer IB type tears originates from the ulnar attachment point of the TFCC, i.e., the base of the ulnar styloid process, along the ligament fiber direction, with a discontinuity extending directly to the articular surface, indicating ligament avulsion from the bone surface. The radially linear high signal intensity path originates from the base of the ulnar styloid process and crosses the palmar articular surfaces of the lunate and triquetrum bones. The severity of this injury type differs only in the length and width of the tear band. Therefore, different injury types, compared to changes in bony relationships, all exhibit regionally limited changes restricted by morphological structures.

[0026] Based on the above analysis, in the embodiments of the present invention, for MRI images, initial region identification is performed to obtain multiple consecutive initial region images. Then, the morphological shift and gradient change of feature points in the initial regions are analyzed in the consecutive initial region images to obtain the variation coefficients of each feature point in the initial regions, wherein: For MRI images, initial region identification is performed to obtain multiple consecutive initial region images. Specifically, this includes acquiring multiple consecutive MRI images from the original dataset, and for each MRI image, using threshold-based segmentation or template matching methods for further fine segmentation to segment specific structural regions such as the ulnar styloid process, lunate bone, triangular bone region, and TFCC as initial regions. The segmented initial regions include the boundaries of each region. For multiple consecutive MRI images, multiple consecutive initial region images are obtained, where multiple consecutive initial region images refer to consecutive slices of the MRI images on the Z-axis.

[0027] Analyze the morphological shift and gradient changes of feature points in the initial region across continuous initial region images to obtain the variation coefficients of each feature point in the initial region, further including: First, feature points are extracted from multiple consecutive initial region images, and then these feature points are matched. Specifically, Harris corner detection is used to extract feature points from multiple consecutive initial region images, obtaining all feature points for each initial region in each initial region image. A feature point tracking / matching algorithm (such as KLT optical flow) is then introduced to track and match the feature points from the multiple consecutive initial region images. Additionally, a coordinate system is established on the initial region image plane to obtain the coordinate data of each feature point.

[0028] Then, for the corresponding structure of each initial region, in addition to being affected by the damage type, it is also affected by changes in the bony relationship. The more the change in the relationship is a cutoff manifestation within a single initial region, the less the damage type of the initial region is affected by the relative structural relationship. The morphological offset and gradient change of the feature points are represented by the coordinate offset and grayscale difference of the edge or texture feature points of the same anatomical structure between adjacent slices (adjacent initial region images), reflecting the continuity or interruption of the tissue in the spatial dimension. The changes in the initial region images obtained from two adjacent slices include the combined effects of damage type manifestation and changes in bony relationship. Therefore, for the matched feature points, by analyzing the grayscale differences of the feature points in the initial region in all adjacent initial region images, the gradient change value of the feature points in each initial region in all adjacent initial region images is obtained; and, for the matched feature points, by analyzing the coordinate point differences of the feature points in the initial region in all adjacent initial region images, the morphological offset value of the feature points in each initial region in all adjacent initial region images is obtained. Furthermore, based on the gradient change value and the shape offset value, the variation coefficient of each feature point in each initial region is obtained. Specifically, the first... The first initial region The formula for calculating the coefficient of variation of each feature point is: In the formula, Indicates the first The first initial region The coefficient of variation of each feature point; Indicates the first The first initial region The gradient change value of each feature point, that is, the difference in the average gray value of all adjacent matching feature points in two initial region images; Indicates the first The first initial region The shape offset value of each feature point, that is, the average coordinate difference of the matching feature points of all two adjacent initial region images; This represents the maximum-minimum normalization function, used to normalize the variation coefficients corresponding to all feature points within the same initial region by taking their maximum and minimum values. The maximum and minimum values ​​in the maximum-minimum normalization function are the values ​​corresponding to the variation coefficients of all feature points within the same initial region. The maximum and minimum values ​​in the range. It should be noted that if... When it is 0, the variation coefficient is directly set. Set to 1.

[0029] The greater the difference in the average gray value of the matching feature points between any two adjacent initial regions, and the smaller the difference in the average coordinates of the matching feature points between any two adjacent initial regions, that is, when the positions are close but the gray values ​​are different, the greater the coefficient of variation of the feature point in the corresponding initial region.

[0030] Similarly, obtain the variation coefficient of each feature point in all initial regions.

[0031] S300: Based on the variation coefficient, analyze the extension performance of the bony structure in the initial region to obtain the regional variation cutoff coefficient of the initial region.

[0032] By obtaining the variation coefficient of each feature point in all initial regions through step S200, each variation coefficient represents the structural change of the local region under region matching. The actual region is affected by the type of injury or by changes in the bony relationship, which is reflected in the correlation of the variation coefficient of the corresponding feature point.

[0033] Correlation refers to the indirect signs that bony structural changes are usually caused by multiple regional structural locations and soft tissue complications. The specific manifestations of the injury type are direct signs such as extended tears or local degeneration within the region. The extended manifestations of these direct signs can effectively reflect the characteristics of the corresponding injury type.

[0034] Based on the above analysis, in an embodiment of the present invention, the extension performance of the bony structure in the initial region is analyzed according to the variation coefficient to obtain the regional variation cutoff coefficient of the initial region. Further steps include: First, based on the coefficients of variation and the coordinate data of feature points, the extension direction of the bony structure in the initial region is obtained. Then, curve fitting and curve segmentation are performed to obtain multiple sub-curves corresponding to each initial region. Specifically, principal component analysis (PCA) is performed for each initial region. This involves using the coefficients of variation and coordinate data of all feature points in each initial region to obtain the first principal component direction as the extension direction of the bony structure in that initial region. Specifically, the coefficients of variation for each feature point in each initial region are normalized by dividing by the sum of the coefficients of variation for all feature points. The normalized value is used as the weight of the coordinate data for each feature point. This weight is then used to perform a weighted multiplication of the coordinate data for each feature point to obtain the weighted coordinate data. PCA is then performed on the weighted coordinate data of all feature points in each initial region. This process yields the extension direction of the bony structure in the initial region.

[0035] The projection of all feature points in the extension direction is a regional variation cutoff feature that restricts the morphological structure of specific signs. This cutoff feature refers to the obvious continuous fluctuation of the variation coefficient in the extension direction, and the significant part of its variation belongs to the influence of direct signs.

[0036] Therefore, using the extension direction as the horizontal axis and the direction perpendicular to the extension direction as the vertical axis, the x-values ​​of the projection coordinates of all feature points in the initial region onto the horizontal axis are used as the horizontal axis coordinates, and the variation coefficients of all feature points in the initial region are used as the vertical axis coordinates. Curve fitting is then performed on all feature points in the initial region to obtain the fitted curve for each initial region. It should be noted that for feature points with the same projection point, their vertical axis coordinates are the average variation coefficients of those feature points. In a specific embodiment, the curve fitting method can be a cubic polynomial curve fitting algorithm (or a cubic spline interpolation algorithm), thereby enabling smooth curve fitting for all feature points in the initial region, establishing a continuously differentiable data model, and obtaining the fitted curve for each initial region.

[0037] Furthermore, regarding the fluctuation of the coefficient of variation of the fitted curve, the curve is segmented according to the slope of all data points on the fitted curve. When the slope of the data point on the fitted curve is greater than the average slope of all data points, it can be used as the segmentation point of the fitted curve, dividing each fitted curve into multiple sub-curves, thus obtaining multiple sub-curves corresponding to each initial region.

[0038] Each sub-curve represents the area of ​​influence of possible direct or indirect signs. For the entire fitted curve, the greater the difference between the interval corresponding to the sub-curve in the middle of the fitted curve and the intervals corresponding to the sub-curves on both sides, the more obvious the regional variation cutoff characteristic of its morphological structure restriction is.

[0039] Therefore, based on the mean coefficient of variation of each sub-curve, cluster analysis is performed on all sub-curves corresponding to each initial region, resulting in 3 clusters. Specifically, the mean coefficient of variation of each sub-curve is obtained, i.e., the mean coefficient of variation of all feature points contained within each sub-curve is calculated. Based on the mean coefficient of variation of each sub-curve, hierarchical cluster analysis is performed on all sub-curves corresponding to each initial region, with a preset number of clusters of 3, resulting in 3 clusters. It should be noted that if the actual number of clusters is less than 3 (e.g., due to a unidirectional data distribution), the mean of the missing set is set to the global mean or 0 to ensure that subsequent procedures can be executed.

[0040] Furthermore, the average variation coefficient of each cluster is obtained, and the differences in the average variation coefficients between clusters are analyzed to obtain the regional variation cutoff coefficient for each initial region. Specifically, the first... The three clusters in the initial region are sorted according to their spatial location and numbered as cluster 1, cluster 2, and cluster 3; the cluster 3 is then obtained. The average variation coefficient of the three clusters in the initial region is calculated, i.e., the average variation coefficient of the entire interval composed of all sub-curves of the three clusters is calculated; then the absolute value of the difference between the average variation coefficients of the second cluster (the cluster in the middle position) and those of the first and third clusters is calculated, and the average variation coefficient of the third cluster is calculated. The average total variation coefficient of the entire interval composed of all sub-curves of the initial region is used to construct the first... The formula for calculating the regional variation cutoff coefficient of an initial region is as follows: In the formula, Indicates the first Regional variation cutoff coefficient for an initial region; Indicates the first The average coefficient of variation of the entire interval consisting of all sub-curves of the first cluster in the initial region; Indicates the first The average coefficient of variation of the entire interval consisting of all sub-curves of the second cluster in the initial region; Indicates the first The average coefficient of variation of the entire interval consisting of all sub-curves of the third cluster in the initial region; Indicates the first The mean of the variation coefficients of the entire interval consisting of all sub-curves of the initial region; Indicates the number of differences between sets; Indicates taking the absolute value; This represents a linear normalization function, such as a max-min normalization function, which normalizes the regional variation cutoff coefficients of all initial regions. The range of values ​​after normalization is... .

[0041] The larger the value, the greater the difference in the average coefficient of variation between the interval corresponding to the sub-curve in the middle of the fitted curve and the interval corresponding to the sub-curves on both sides, which further proves that the regional variation cutoff characteristic of its morphological structure restriction is more obvious.

[0042] S400: Perform convolutional analysis on the initial region to obtain the feature maps of each convolutional layer. Combine the regional variation cutoff coefficient to construct the regional feature coefficient map of the initial region. Then analyze the significant performance of the initial region in each convolutional layer to obtain the attention performance of the initial region in each convolutional layer.

[0043] TFCC damage (such as tears) manifests as specific morphological discontinuities (i.e., cutoff appearance) on spatial slices. Conventional convolutional networks tend to smooth out these subtle features. This invention improves recognition accuracy by explicitly calculating regional variation cutoff coefficients as prior knowledge to guide the network to focus on these abrupt changes. Since the obtained regional variation cutoff coefficients of the initial region are primarily based on the feature representation of the TFCC damage type contained in the corresponding initial region, analyzed from the actual MRI image variation structure, this is merely a simplified classification based on the acquired MRI image signs. The actual recognition result of the damage type mainly requires analysis of the preservation of this feature within the multiple convolutional layers of the deep neural network; that is, further analysis of the attention feature representation of the regional variation cutoff coefficients of the initial region at multiple levels is needed.

[0044] The U-Net fully convolutional neural network in deep learning neural networks, which significantly impacts damage recognition results, mainly consists of three parts: the encoder, the bottleneck layer, and the decoder. It achieves multi-scale feature fusion through skip connections. The role of skip connections is to directly pass shallow, high-resolution features from the encoder—preserving edge and texture details—to the corresponding layers in the decoder, compensating for spatial information loss caused by downsampling. Considering the regional variation cutoff coefficient of the initial region involves analyzing the correlation between feature loss during convolution and these skip connections, aiming to retain as many variation features of the initial region as possible for accurate damage type identification.

[0045] Based on the above analysis, in the embodiments of the present invention, by performing convolutional analysis on the initial region to obtain the feature maps of each convolutional layer, and combining the regional variation cutoff coefficient, a regional feature coefficient map of the initial region is constructed, further including: First, the initial region is analyzed by using the U-Net fully convolutional neural network in deep learning neural networks to obtain the feature map after multiple samplings during the convolution process, that is, to obtain the feature map of each initial region in each convolutional layer.

[0046] Then, the regional variation cutoff coefficient of the initial region is regarded as the feature coefficient of all pixels in the initial region, and a regional feature coefficient map of the same size as the feature map is constructed for each initial region.

[0047] For different damage types, it is desirable to show different significant features on different convolutional layers. This proves that the more accurate the corresponding feature recognition results are, the more specific the regional variation cutoff coefficients of different initial regions are on different convolutional layers, showing clustering characteristics and uniform distribution characteristics overall. Therefore, the more uniform the abnormal performance of the regional feature coefficient maps of all initial regions on multiple convolutional layers, the more significant the attention performance of the corresponding deep convolutional layers.

[0048] Therefore, by analyzing the salient performance of the initial region in each convolutional layer, the attention performance of the initial region in each convolutional layer is obtained, further including: First, the differences between the initial region's feature coefficient map and the feature maps of the initial region in each convolutional layer are analyzed to obtain the abnormal performance of the region's feature coefficient map in each convolutional layer. Specifically, the 1st... The regional characteristic coefficient map of the initial region and the first... The initial region in the first The absolute value of the difference between the feature maps of the nth convolutional layer, combined with the number of convolutional layers, is used to construct the nth convolutional layer. The regional characteristic coefficient map of the initial region is in the first... The formula for calculating the abnormal behavior on each convolutional layer is: In the formula, Indicates the first The regional characteristic coefficient map of the initial region is in the first... Abnormal behavior on each convolutional layer; Indicates the first The initial region in the first Feature maps of each convolutional layer; Indicates the first Regional characteristic coefficient diagram of the initial region; Indicates the first The number of convolutional layers in each initial region; Indicates taking the absolute value; This represents the batch normalization function, which performs batch normalization on the output of each channel of the feature map separately. The normalized values ​​range from [value range missing]. .

[0049] The larger the value, the more significant the effect. The initial region in the first Different salient features are exhibited on each convolutional layer, namely the th... The initial region in the first The specificity of the convolutional layer exhibits clustering characteristics, corresponding to the th convolutional layer. The regional characteristic coefficient map of the initial region is in the first... The more pronounced the anomalies on each convolutional layer, the more significant they become.

[0050] Then, based on the abnormal behavior, the differences in abnormal behavior of the initial region among all adjacent convolutional layers are analyzed, as well as the differences between the abnormal behavior of the region feature coefficient map of the initial region on each convolutional layer and the mean of abnormal behavior on all convolutional layers are analyzed, thus obtaining the significant behavior of the initial region on each convolutional layer. Specifically, the calculation of the first... The absolute value of the anomalous performance difference between the initial region and all adjacent convolutional layers, and the calculation of the first... The regional characteristic coefficient map of the initial region is in the first... The absolute value of the difference between the outlier performance on the nth convolutional layer and the mean outlier performance across all convolutional layers is used to construct the nth convolutional layer. The initial region in the first The formula for calculating the salient performance of the nth convolutional layer, i.e., the nth convolutional layer The initial region in the first The formula for calculating the attention performance of each convolutional layer is: In the formula, Indicates the first The initial region in the first The significant performance of the first convolutional layer, namely the first... The initial region in the first Attention performance of each convolutional layer; Indicates the first The regional characteristic coefficient map of the initial region is in the first... Abnormal behavior on each convolutional layer; Indicates the first The regional characteristic coefficient map of the initial region is in the first... Abnormal behavior on each convolutional layer; Indicates the first Number of convolutional layers in each initial region; Indicates the first The mean of the anomalous performance of the region feature coefficient map of the initial region across all convolutional layers; Indicates taking the absolute value; This represents the denominator correction parameter, which has the same dimensions as the denominator and takes the smallest value greater than 0. This is to prevent the denominator from being 0; for example, it can be set... This can prevent the denominator from being 0, and will not significantly interfere with the calculation of normal and abnormal values. Represented by natural constant An exponential function with base 0.

[0051] pass To measure and If the changes in both are consistent, and the consistency between their changes is relatively large, then the ratio between the two is... The closer to 1, that is The smaller the value, the more uniform the distribution of anomalies across all convolutional layers, indicating a greater consistency in their variation. This suggests that the current anomaly corresponds to significant feature extraction across different convolutional layers, meaning that the attention-based feature extraction across multiple convolutional layers is highly effective. Therefore, through... Measure the first The initial region in the first The smaller the value, the more uniform the abnormal performance across all convolutional layers, and the more significant the attention performance of the initial region in that convolutional layer.

[0052] S500: Based on the attention performance, analyze the minimum difference between the initial region and other initial regions, adjust the attention weights of the initial region in each convolutional layer, and then train the convolutional neural network to achieve TFCC damage recognition.

[0053] The attention performance of all initial regions in each convolutional layer reflects the feature analysis during the training of the U-Net fully convolutional neural network. Therefore, a spatial attention module needs to be introduced at the encoder-decoder level of the U-Net during actual training. This spatial attention module dynamically calculates the importance weight of each pixel position, and its weight analysis depends on the actual distribution of attention performance in the overall initial region image. Training the U-Net fully convolutional neural network based on the adjusted attention weights can effectively improve the accuracy of TFCC lesion identification in wrist MRI.

[0054] Based on the above analysis, in the embodiments of the present invention, based on the attention performance, the minimum difference between the initial region and other initial regions is analyzed, and the attention weights of the initial region in each convolutional layer are adjusted, thereby training the convolutional neural network to achieve TFCC damage recognition. Wherein: Based on attention performance, the minimum difference between the initial region and other initial regions is analyzed, and the attention weights of the initial region in each convolutional layer are adjusted. Further adjustments include: Based on the actual distribution of attention performance across the entire initial region image, the attention weights of the current convolutional layer are adjusted. Each convolutional layer has an attention weight matrix of the same size as the feature map. For each convolutional layer, if the attention performance of any initial region in the entire initial region image shows a small difference from the other initial regions, it indicates that the feature extraction performance of that convolutional layer is poor, and its adjustment process needs to set the attention weights based on this minimum difference. Therefore, firstly, for each convolutional layer, the minimum absolute value of the difference between the attention performance of the initial region and the attention performance of all other initial regions is obtained, thus determining the minimum difference between the initial region and the other initial regions. Then, based on this minimum difference, an adjustment coefficient and an inverse proportional exponential function are introduced to limit the adjustment range within a reasonable range (e.g., ...). To prevent the weights from exploding, the attention weights of the initial region in each convolutional layer are adjusted, resulting in adjusted attention weights for the initial region in each convolutional layer. Specifically, the first... The initial region in the first The formula for calculating the adjusted attention weights of each convolutional layer is: In the formula, Indicates the first The initial region in the first Adjusting the attention weights of each convolutional layer; Indicates the first The initial region in the first The initial attention weights of each convolutional layer, i.e. the coefficient values ​​of the spatial attention map at the corresponding positions, are obtained through the U-Net fully convolutional neural network; Indicates the first The initial region in the first Attention performance of each convolutional layer; Indicates the first The first region outside the initial region The initial region in the first The attention performance of each convolutional layer, among which Indicates the first The number of initial regions outside the initial regions; Represented by natural constant An exponential function with base 0 is used for inverse proportional normalization. This represents the adjustment coefficient, used to control the intensity of attention adjustment, and its value can be 0.1; This represents the function that takes the minimum value.

[0055] Similarly, the attention weights of all initial regions in each convolutional layer are adjusted.

[0056] Training a convolutional neural network to achieve TFCC damage recognition, further including: First, by adjusting the attention weights of all initial regions in each convolutional layer, we obtain the attention weight matrix for each convolutional layer corresponding to the feature map size. Specifically, we adjust the attention weights of all initial regions in the first convolutional layer. The attention weights of each convolutional layer are adjusted and concatenated to form a complete feature map (the first one). The attention weight matrix of size (the feature map of the nth convolutional layer) is obtained to obtain the nth convolutional layer's feature map. Attention weight matrix for each convolutional layer.

[0057] Then, the attention weight matrices of each convolutional layer are used for training the U-Net fully convolutional neural network to obtain a trained U-Net fully convolutional neural network, thereby realizing the training of the U-Net fully convolutional neural network with attention weights.

[0058] Then, the MRI images are input into the trained U-Net fully convolutional neural network, and the output is a multi-channel probability map, with each channel corresponding to a type of injury. Specifically, in actual use, the wrist joint MRI images (PD-FS or T2WI sequences) to be analyzed are normalized and input into the trained U-Net fully convolutional neural network model. The model outputs a multi-channel probability map, with each channel corresponding to a type of injury, such as IA, IB, II-A, and II-B.

[0059] Finally, threshold segmentation (e.g., a segmentation threshold of 0.5) is performed on the multi-channel probability map to generate a binarized mask, which is then used to mark the pixel positions of each damaged area on the display.

[0060] This embodiment also provides a deep learning-based TFCC lesion recognition system in wrist MRI, including: a memory and a processor; the memory is connected to the processor, and the memory is used to store program instructions; the processor is used to implement the steps in the above-mentioned deep learning-based TFCC lesion recognition method in wrist MRI when the program instructions are executed.

Claims

1. A deep learning-based method for identifying TFCC lesions in wrist joint MRI, characterized in that, The method includes: Obtain MRI images of the patient's wrist joint; For the MRI image, initial region identification is performed to obtain multiple consecutive initial region images. Then, the morphological shift and gradient change of feature points in the initial region in the consecutive initial region images are analyzed to obtain the variation coefficient of each feature point in the initial region. Based on the variation coefficient, the extension performance of the bony structure in the initial region is analyzed to obtain the regional variation cutoff coefficient of the initial region. Convolutional analysis is performed on the initial region to obtain the feature maps of each convolutional layer. Combined with the regional variation cutoff coefficient, a regional feature coefficient map of the initial region is constructed. Then, the significant performance of the initial region in each convolutional layer is analyzed to obtain the attention performance of the initial region in each convolutional layer. Based on the attention performance, the minimum difference between the initial region and other initial regions is analyzed, the attention weights of the initial region in each convolutional layer are adjusted, and then a convolutional neural network is trained to achieve TFCC damage recognition.

2. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 1, characterized in that, For the MRI image, initial region identification is performed to obtain multiple consecutive initial region images, including: For the MRI image, threshold segmentation is performed to identify the initial regions corresponding to the lunate bone, triquetrum bone, styloid process of the ulna, and TFCC of the wrist joint, and multiple continuous initial region images are obtained, where multiple continuous initial region images refer to continuous slices of the MRI image on the Z-axis.

3. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 2, characterized in that, Analyze the morphological shift and gradient change of feature points in the initial region in the continuous initial region images to obtain the variation coefficients of each feature point in the initial region, including: Feature points are extracted from multiple consecutive initial region images, and the feature points in the multiple consecutive initial region images are matched. For the matched feature points, the grayscale differences of the feature points in the initial region are analyzed in all adjacent initial region images to obtain the gradient change value of each feature point in the initial region in all adjacent initial region images. For the matched feature points, the differences in coordinate points of the feature points in the initial region among all adjacent initial region images are analyzed to obtain the morphological offset value of each feature point in the initial region among all adjacent initial region images. Based on the gradient change value and the shape offset value, the variation coefficient of each feature point in each initial region is obtained.

4. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 1, characterized in that, Based on the variation coefficient, the extension behavior of the bony structure in the initial region is analyzed to obtain the regional variation cutoff coefficient of the initial region, including: Based on the variation coefficient and the coordinate data of the feature points, the extension direction of the bony structure of the initial region is obtained, and then curve fitting and curve segmentation are performed to obtain multiple sub-curves corresponding to each initial region. Based on the mean variation coefficient of each sub-curve, cluster analysis is performed on all the sub-curves corresponding to each initial region to obtain 3 clusters; The average variation coefficient of each cluster is obtained, and the differences in the average variation coefficients between clusters are analyzed to obtain the regional variation cutoff coefficient of each initial region.

5. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 4, characterized in that, Based on the variation coefficient and the coordinate data of the feature points, the extension direction of the bony structure in the initial region is obtained, and then curve fitting and curve segmentation are performed to obtain multiple sub-curves corresponding to the initial region, including: Based on the variation coefficient and the coordinate data of the feature points, PCA principal component analysis is performed to obtain the direction of the first principal component as the extension direction of the bony structure of the initial region; Using the extension direction as the horizontal axis and the direction perpendicular to the extension direction as the vertical axis, the x-values ​​of the projection coordinates of all feature points in the initial region on the horizontal axis are used as the horizontal axis coordinates, and the variation coefficients of all feature points in the initial region are used as the vertical axis coordinates. Curve fitting is performed on all feature points in the initial region to obtain the fitting curve for each initial region. The curve is segmented based on the slope of all data points on the fitted curve to obtain multiple sub-curves corresponding to each initial region.

6. The method for identifying TFCC lesions in wrist joint MRI based on deep learning according to claim 1, characterized in that, Convolutional analysis is performed on the initial region to obtain the feature maps of each convolutional layer. Combined with the regional variation cutoff coefficient, a regional feature coefficient map of the initial region is constructed, including: The initial region is analyzed by convolutional analysis using the U-Net fully convolutional neural network to obtain the feature map of each initial region in each convolutional layer; The regional variation cutoff coefficient of the initial region is regarded as the feature coefficient of all position pixels in the initial region, and a regional feature coefficient map of the same size as the feature map is constructed for each initial region.

7. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 6, characterized in that, The analysis of the significant performance of the initial region in each convolutional layer includes: By analyzing the differences between the region feature coefficient map of the initial region and the feature map of the initial region in each convolutional layer, the abnormal performance of the region feature coefficient map of each initial region in each convolutional layer is obtained. Based on the abnormal behavior, the differences in abnormal behavior of the initial region among all adjacent convolutional layers are analyzed, and the differences between the abnormal behavior of the region feature coefficient map of the initial region on each convolutional layer and the mean of abnormal behavior on all convolutional layers are analyzed to obtain the significant behavior of the initial region on each convolutional layer.

8. The method for identifying TFCC lesions in wrist joint MRI based on deep learning according to claim 1, characterized in that, Based on the attention performance, the minimum difference between the initial region and other initial regions is analyzed, and the attention weights of the initial region in each convolutional layer are adjusted, including: For each convolutional layer, the minimum absolute value of the difference between the attention performance of the initial region and the attention performance of all other initial regions is obtained, thus obtaining the minimum difference between the initial region and the other initial regions. Based on the minimum difference, and by introducing an adjustment coefficient and an inverse proportional exponential function, the attention weights of the initial region in each convolutional layer are adjusted to obtain the adjusted attention weights of the initial region in each convolutional layer.

9. The deep learning-based TFCC lesion identification method in wrist joint MRI according to claim 8, characterized in that, Training a convolutional neural network to achieve TFCC damage recognition includes: Based on the adjusted attention weights of all the initial regions in each convolutional layer, an attention weight matrix corresponding to the feature map size is obtained for each convolutional layer; The attention weight matrices of each convolutional layer are used to train the U-Net fully convolutional neural network to obtain the trained U-Net fully convolutional neural network. The MRI image is input into a trained U-Net fully convolutional neural network, and the output is a multi-channel probability map, with each channel corresponding to a type of damage; The multi-channel probability map is thresholded to generate a binary mask, which identifies the pixel position of each damaged area.

10. The deep learning-based method for identifying TFCC lesions in wrist joint MRI according to claim 1, characterized in that, Obtain MRI images of the patient's wrist joint, including: Initial MRI images of the patient's wrist joint, including proton density-weighted fat-suppressed sequences and T2-weighted sequences, were acquired. For the initial MRI images, the proton density-weighted fat-suppressed sequence and the T2-weighted image sequence were rigidly registered to obtain the registered MRI images; For the registered MRI images, a three-dimensional local region is cropped with the distal ulnar head of the wrist joint as the center, focusing on the TFCC and the lunate, triquetrum, and ulnar styloid processes, and removing irrelevant tissues to obtain the MRI images of the patient's wrist joint.

Images