A diagnostic method and system for glenoid and humeral head defect area

Abstract

The application discloses a kind of diagnostic methods and systems for glenoid and humeral head defect area, belong to image processing technical field, the method includes: obtaining the medical image data of the glenoid and humeral head to be processed;The medical image data is input into the segmentation model of pre-setting, and the segmentation result of glenoid and humeral head is obtained;Based on the segmentation result, the defect area and defect proportion of glenoid are calculated;Based on the defect proportion, the defect degree of humeral head is determined.The diagnostic method and system for glenoid and humeral head defect area of the application provide a kind of efficient, accurate intelligent diagnostic auxiliary tool for clinic, with good application prospect and popularization value.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a diagnostic method and system for defects in the glenoid cavity and humeral head. Background Technology

[0002] Shoulder dislocation is a common traumatic condition in clinical practice, often resulting in bone damage to the humeral head and glenoid cavity. Humeral head defects and glenoid defects are two common pathological conditions of the shoulder joint, severely affecting shoulder joint stability and function, significantly impacting patients' daily lives and athletic abilities. With increasing exercise intensity and the occurrence of traumatic accidents, the incidence of humeral head and glenoid defects is rising annually. Humeral head defects mainly manifest as localized bone damage or absence, while glenoid defects present as partial or complete absence of the glenoid cavity. These defects can lead to a series of subsequent problems such as joint degeneration and arthritis. Therefore, early detection and accurate assessment of the area of ​​these defects can provide important information for clinical treatment and intervention.

[0003] Traditional diagnostic methods for humeral head and glenoid defects rely on X-rays, CT scans, and MRI images, with the extent and severity of the damage determined by professional physicians through manual analysis. However, these methods are susceptible to interference from physician experience, image noise, and complex morphology, resulting in low diagnostic accuracy and efficiency. This is especially true when processing large volumes of images, where physician judgment may be biased, affecting treatment decisions. Summary of the Invention

[0004] This invention provides a diagnostic method and system for assessing the area of ​​defects in the glenoid fossa and humeral head, addressing the technical problem that traditional diagnostic methods for humeral head and glenoid defects rely on manual analysis to determine the extent and severity of the damage. This results in low diagnostic accuracy and efficiency, impacting treatment decisions.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] On one hand, the present invention provides a diagnostic method for the area of ​​defects in the glenoid cavity and humeral head, the diagnostic method for the area of ​​defects in the glenoid cavity and humeral head including:

[0007] Acquire medical imaging data of the glenoid cavity and humeral head to be processed;

[0008] The medical image data is input into a preset segmentation model to obtain the segmentation results of the glenoid cavity and the humeral head;

[0009] Based on the segmentation results, the defect area and defect ratio of the glenoid cavity are calculated;

[0010] Based on the aforementioned defect ratio, the degree of humeral head defect is determined.

[0011] Furthermore, the segmentation model includes multiple feature enhancement modules constituting a Transformer structure; starting from the second feature enhancement module, the output of each feature enhancement module is processed sequentially by an adaptive medical convolution module and an edge attention mechanism module, and the resulting feature data is input into the segmentation head; the segmentation head, based on the input multiple sets of feature data, achieves segmentation of the glenoid cavity and the humeral head, and outputs the segmentation results of the glenoid cavity and the humeral head.

[0012] Furthermore, the feature enhancement module includes an overlapping block embedding module and multiple feature enhancement units; wherein, the multiple feature enhancement units are connected in series, and adjacent feature enhancement units are connected through a downsampling layer;

[0013] The overlapping block embedding module introduces overlapping patch coding, transforming the input image through a 7×7 convolutional layer. Each vector in the result represents an embedded overlapping region with context, serving as the token sequence input to the Transformer structure. The input sequence x to the feature enhancement unit first passes through a feedforward network to obtain the output x′, expressed by the formula:

[0014] x′=x+γ·Conv 1×1 (LN(x))+FFN(x)

[0015] Where γ is a learnable scaling parameter; Conv 1×1 The local convolutional feedback channel is used to embed spatial structure information; LN(x) represents normalization of the input; FFN represents independent transformation of each token, and its transformation operation is expressed as:

[0016] FFN(x)=W2·σ(W1x+b1)+b2

[0017] Where W1 and W2 represent weight matrices, σ represents a nonlinear activation function used to normalize the input, and b1 and b2 represent bias vectors;

[0018] x′ then enters the dimensional interaction module to enhance the information exchange between tokens at different locations; then, through the cascaded group attention mechanism module, x′ processed by the dimensional interaction module is divided into G groups in the channel dimension, and local self-attention is performed independently on each group to reduce the amount of computation.

[0019] Finally, the outputs of each group are concatenated and then subjected to information fusion, which is used as the output of the feature enhancement unit.

[0020] Furthermore, the number of feature enhancement units is 3; G is set to 16 in the first feature enhancement unit, 8 in the second feature enhancement unit, and 4 in the last feature enhancement unit.

[0021] Furthermore, the operational formula for the local self-attention is expressed as follows:

[0022]

[0023] Among them, Q (i) K represents the i-th Query matrix; (i) V is the i-th key matrix; (i) Let d be the i-th Value matrix; k The dimension of the key; T represents the transpose of the matrix; Z (i) This represents the output of the i-th attention head; Softmax() represents the normalization process.

[0024] Furthermore, the data processing procedure of the edge attention mechanism module includes:

[0025] The Sobel operator is used to extract the first-order gradient information of the input image in the horizontal and vertical directions, respectively; where, the horizontal operator G... x and vertical direction operator G y Represented as:

[0026]

[0027] Each channel of the input image is respectively compared with G x Perform a convolution operation to obtain the horizontal gradient map, and then convolve each channel of the input image with G. y Perform a convolution operation to obtain the gradient map in the vertical direction;

[0028] The gradient maps in the horizontal and vertical directions are processed by a global average pooling layer to obtain the aggregate tensor T of the input image in the vertical direction. H and the aggregation tensor T in the horizontal direction W ;

[0029] T H and T W Multiplying them together yields a location-dependent attention map M;

[0030] M is input into a 1×1 convolutional layer, and then the output of the 1×1 convolutional layer is added to the input image of the edge attention mechanism module to obtain the optimized feature map.

[0031] Furthermore, the data processing procedure for the segmentation head includes:

[0032] The multiple sets of input feature data are fused to obtain the fused feature data F. fused ;

[0033] The weight of each pixel is dynamically calculated; the weight generation formula is as follows:

[0034] W=σ(Conv(F fused ))

[0035] Where W is a dynamically generated weight mask; σ represents a non-linear activation function used to normalize the input; Conv(·) represents a convolution operation; the range of W is [0,1], and the larger the value, the more the model focuses on the corresponding region;

[0036] Based on W, a final dynamic mask is generated to guide segmentation; the formula for generating the dynamic mask is:

[0037] M dynamic =W⊙P

[0038] Among them, M dynamic For dynamic mask; P represents the initial prediction result of the model; ⊙ indicates pixel-wise weighting;

[0039] Based on the generated dynamic mask, segmentation of the glenoid cavity and humeral head is achieved.

[0040] Further, based on the segmentation results, the defect area and defect ratio of the glenoid cavity are calculated, including:

[0041] Using the segmentation model, a standard complete shape of the glenoid cavity is generated based on the training data of healthy samples;

[0042] Using the segmentation model, the actual shape of the glenoid cavity is generated based on the medical imaging data;

[0043] The pixel area difference between the standard complete shape and the actual shape is calculated to obtain the defect area of ​​the glenoid cavity;

[0044] Calculate the percentage of the defect area to the intact glenoid cavity area to obtain the defect ratio of the glenoid cavity.

[0045] Furthermore, based on the stated defect ratio, the degree of humeral head defect is determined, including:

[0046] If the defect rate is less than 10%, it is determined to be a minor defect;

[0047] If the defect rate is between 10% and 30%, it is determined to be a moderate defect;

[0048] If the defect rate is greater than 30%, it is judged as a severe defect.

[0049] On the other hand, the present invention also provides a diagnostic system for the area of ​​defects in the glenoid cavity and humeral head, the diagnostic system for the area of ​​defects in the glenoid cavity and humeral head comprising:

[0050] The data acquisition module is used to acquire medical imaging data of the glenoid cavity and humeral head to be processed;

[0051] The target segmentation module is used to input the medical image data into a preset segmentation model to obtain the segmentation results of the glenoid cavity and the humeral head;

[0052] The data calculation module is used to calculate the defect area and defect ratio of the glenoid cavity based on the segmentation results;

[0053] The defect severity assessment module is used to determine the degree of humeral head defect based on the defect ratio.

[0054] In another aspect, the present invention also provides an electronic device comprising a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method.

[0055] In another aspect, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the above method.

[0056] The beneficial effects of the technical solution provided by this invention include at least the following:

[0057] This invention introduces an edge attention mechanism module and a feature fusion segmentation head to address the complex morphology of defects, thereby optimizing the segmentation accuracy of the humeral head region. A feature enhancement module improves feature extraction efficiency and addresses the issue of morphological complexity. A collaborative task-driven mechanism achieves joint optimization of dual-damage identification tasks, improving the overall accuracy and robustness of the model in target recognition and area calculation. Experimental results show that the proposed network exhibits excellent performance in the automatic detection and quantitative analysis of humeral head and glenoid defects, particularly demonstrating good generalization ability in complex background interference and small-area lesion identification. This provides a highly efficient and accurate intelligent diagnostic aid for clinical practice, with promising application prospects and widespread application value. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart of a diagnostic method for defects in the glenoid cavity and humeral head provided in an embodiment of the present invention.

[0060] Figure 2This is a schematic diagram of the DDDNet model structure provided in an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the EGAM module provided in an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the OptiViT network module structure provided in an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of the structure of the ASFFHead segmentation head provided in an embodiment of the present invention;

[0064] Figure 6 This is a schematic diagram of a label example provided in an embodiment of the present invention;

[0065] Figure 7 These are segmentation effect diagrams of the glenoid cavity provided in the embodiments of the present invention; wherein, the first column is the MRI image of the glenoid cavity, the second column is the image annotated by experts, and the rest are segmentation effect diagrams of various segmentation networks in turn;

[0066] Figure 8 These are segmentation results of the humeral head provided in an embodiment of the present invention; wherein, the first column is the MRI image of the humeral head, the second column is the image annotated by experts, and the remaining columns are segmentation results of various segmentation networks in turn;

[0067] Figure 9 This is a structural diagram of the adaptive medical convolution module provided in an embodiment of the present invention;

[0068] Figure 10 This is a system block diagram of the electronic device provided in the embodiments of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0070] First, it should be noted that in the embodiments of the present invention, the words "exemplarily," "for example," etc., are used to indicate that they are examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the term "exemplarily" is intended to present the concept in a specific manner. Furthermore, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either one or the other.

[0071] First Embodiment

[0072] This embodiment provides a diagnostic method for the area of ​​defects in the glenoid cavity and humeral head. This method can be implemented by an electronic device, which can be a terminal or a server.

[0073] The execution flow of this method is as follows: Figure 1 As shown, it includes the following steps:

[0074] S1, acquire medical imaging data of the glenoid cavity and humeral head to be processed;

[0075] S2, input medical imaging data into the preset segmentation model to obtain the segmentation results of the glenoid cavity and humeral head;

[0076] Specifically, the structure of the segmentation model DDDNet in this embodiment is as follows: Figure 2 As shown, it consists of a Transformer-based Feature Enhancement Module (Optimized Vision Transformer-based Feature Enhancement Module, OptiViT). It enhances high-order features at each stage through an adaptive medical convolution module, and uses an Edge-global Attention Module (EGAM) to utilize shallow detail features to improve edge segmentation. In addition, it uses an Adaptively Spatial Feature Fusion Detection Head (ASFFHead) to fuse multi-scale feature maps, further improving the performance of DDDNet.

[0077] Compared to traditional segmentation models, DDDNet effectively maintains high-precision segmentation when handling targets and complex boundaries in glenoid and humeral head images. Through its lightweight design and efficient inference process, DDDNet meets the real-time requirements of medical diagnosis, helping doctors obtain accurate diagnostic results in a shorter time. With its precise segmentation capabilities and efficient real-time processing, DDDNet improves the diagnostic accuracy of orthopedic diseases, demonstrating outstanding accuracy and efficiency in glenoid and humeral head defect segmentation tasks.

[0078] The main functional modules in DDDNet will be described in detail below.

[0079] 1. Edge attention mechanism

[0080] EGAM is an innovative module that combines edge information attention mechanisms, with the following specific structure: Figure 3As shown, its purpose is to enhance the ability of deep learning models to focus on edge regions in visual tasks. Subtle defects and critical edge regions of the glenoid fossa and humeral head have significant diagnostic value in medical image analysis and target segmentation tasks. By combining and fine-tuning global attention and edge attention mechanisms, EGAM can effectively enhance the model's sensitivity to edge regions of humeral head defects, improving detection and segmentation accuracy.

[0081] EGAM enhances the sensitivity of deep learning models to key regions by combining edge information and global features, such as Figure 3 As shown, C, H, and W represent the width, height, and number of channels of the input image, respectively. Edge regions in images often contain important structural information, and in medical images, the contours and shapes of defects are often located in edge regions. EGAM extracts edge features, enabling the model to focus on these detailed regions. First, the Sobel operator is used to extract the first-order gradient information of the image in the horizontal and vertical directions, respectively. Its horizontal operator G... x and vertical direction operator G y Represented as:

[0082]

[0083] For a given input image Each channel is connected to G. x G y Perform a convolution operation to obtain gradient maps in the horizontal and vertical directions, then concatenate the two to generate an edge map. Edge maps highlight edge regions in an image. Edge information is encoded into a salient feature map through convolutional neural network layers, providing contextual information about the edges for subsequent steps.

[0084] EGAM utilizes global contextual information to capture important regions in an image and calculates channel-level attention weights based on global features. Spatial average pooling is performed on each channel of the input feature map F to obtain the channel's global contextual information. The average pooling operation is shown in the formula:

[0085]

[0086] Among them, F c,i,j This represents the pixel value of the c-th channel in the i-th row and j-th column of the input feature map.

[0087] Next, the module calculates the aggregate tensor T of the input feature map in the vertical and horizontal directions using a global average pooling layer. H and T WThese two tensors converge spatially to the pixel-level information of the feature map, forming a spatially based weight distribution. These weights represent the importance of different locations, helping to highlight edge regions and detailed information, thereby enhancing the model's sensitivity to edges. Then, T... H and T W Multiplication yields a location-dependent attention map M, which further guides the network to focus on edge regions in the image. Finally, M is passed through a 1×1 convolutional layer and added to the input feature map to obtain the optimized feature map. This process effectively improves the segmentation of highly complex and low-contrast edge regions in the image by enhancing the feature response of edge regions, enabling better reconstruction of details and improving segmentation accuracy and robustness.

[0088] 2. OptiViT module

[0089] Medical images typically have high resolution and a lot of redundant information in segmentation tasks. Directly using standard convolution is not only computationally expensive but may also lead to over-smoothing of features. To accelerate the extraction of damage features, this embodiment designs the OptiViT network as the network, and to better extract local information in glenoid fossa injury images, the specific network structure diagram is shown below. Figure 4 As shown.

[0090] The defects have complex morphologies, exhibiting anatomical features characterized by both local continuity and irregularity in the defect area. Overlapping patch coding is introduced to encode the input image... A 7×7 convolutional layer transformation is performed, and each vector in the result represents an embedded overlapping region with context, which serves as the token sequence input to the Transformer. The input sequence x to the OptiVit unit first passes through a feedforward feedback network, specifically as shown in Equation (1), to obtain the output x′.

[0091] x′=x+γ·Conv 1×1 (LN(x))+FFN(x) (1)

[0092] Where γ is a learnable scaling parameter, Conv 1×1 For local convolutional feedback channels, spatial structure information is embedded. LN(x) represents normalization of the input, and FFN represents independent transformation of each token. The transformation operation is given by formula (2).

[0093] FFN(x)=W2·σ(W1x+b1)+b2 (2)

[0094] Here, W1 and W2 represent weight matrices, σ represents a non-linear activation function used to normalize the input, and b1 and b2 represent bias vectors, which are learnable translation parameters. Then, Token Interaction (a mechanism inspired by the Token-to-Token ViT model, which achieves explicit interaction modeling between tokens through multi-step token transformation, effectively improving global modeling capabilities) is introduced to enhance information exchange between tokens at different positions. A cascaded grouping attention mechanism module is embedded in the Transformer structure. This module divides the input feature image into G groups along the channel dimension to obtain features. Figure X .

[0095]

[0096] In the first OptiVit cell, G is set to 16; in the second OptiVit cell, G is set to 8; and in the last OptiVit cell, G is set to 4.

[0097] Perform local self-attention independently for each group, reducing computational cost. The formula for local attention operation is as follows:

[0098]

[0099] Among them, Q (i) Let K represent the i-th Query matrix. (i) Let V be the i-th key matrix. (i) Let d be the i-th Value matrix. k The dimension of the key; T represents the transpose of the matrix; Z (i) Let Z represent the output matrix of the i-th attention head, which integrates information from the current query and all key-value pairs. Softmax() normalizes the attention score, converting it into a probability distribution to highlight the weights of key matches and emphasize important term pairs. The outputs of each group are concatenated and then fused to produce Z = Concat(Zi). (1) ,...,Z (G)Subsequently, information from other groups is integrated to enhance global expressive power. After adding the OptiViT Block to the backbone network, the model can extract glenoid features from both local and global perspectives. On the one hand, by capturing small-scale details, it accurately separates the boundaries between complete and defective regions. On the other hand, through parallel processing of multiple convolutions, the model can effectively distinguish significant differences between normal and abnormal glenoid features by focusing on the overall morphological characteristics of the glenoid. OptiViT enhances the high-resolution representation of glenoid edges, arc shapes, and defective regions in the feature extraction stage of the segmentation model, and can still output stable segmentation results even under conditions of high image noise or incomplete glenoid morphology.

[0100] 3. ASFFHead Segmentation Header

[0101] ASFFHead plays a crucial role in defect area detection tasks. The humeral head and glenoid cavity may suffer damage of varying degrees, and accurate defect area calculation and precise defect segmentation rely on the effective fusion of multi-scale information. For example... Figure 5 As shown, ASFFHead improves the detection accuracy of humeral head defects by optimizing the model's response to features at different scales through its feature aggregation and selective fusion mechanisms when dealing with humeral head defects of different sizes and shapes.

[0102] Noise and complex backgrounds are common problems in medical imaging, blurring the boundaries of humeral head and glenoid defects. ASFFHead's selective fusion mechanism allows the network to dynamically select the most relevant features to fuse based on the importance of the defect region. In this way, the model can focus on the defect portion of the humeral head and glenoid against complex backgrounds, balancing global and local features to improve defect detection accuracy. Since the humeral head is typically approximately circular, ASFFHead helps the model effectively capture this morphological regularity and accurately segment and identify regions based on bone defects and fractures in the image.

[0103] Shallow features reflect boundary contours and texture information, while deep features focus on high-level semantic expression. The two are integrated through a feature fusion module to achieve a comprehensive representation of the defect region. The feature fusion formula is:

[0104] F fused =Concat(F shallow ,F deep (5)

[0105] Among them, F shallow and F deepThese represent features extracted from the shallow and deep networks, respectively, with Concat(*) indicating a feature concatenation operation. The key to dynamic mask generation lies in adjusting the weights of the regions of interest based on the specific characteristics of the input image. By constructing a weight generation network, the weight of each pixel is dynamically calculated, thereby highlighting salient areas such as damaged regions and glenoid morphology. The weight generation formula is:

[0106] W=σ(Conv(F fused (6)

[0107] Where W is the dynamically generated weight mask, σ represents the non-linear activation function used to normalize the input, and Conv(·) represents the convolution operation. The weight W ranges from [0,1], with larger values ​​indicating greater model focus on that region. Based on the dynamically generated weight W, the module generates the final dynamic mask to guide segmentation. The dynamic mask generation formula is:

[0108] M dynamic =W⊙P (7)

[0109] Among them, M dynamic Here, P represents the initial prediction result of the model, and ⊙ denotes pixel-wise weighting. Multi-scale objects are typically handled in traditional object detection models using pyramid structures or shallow-to-deep feature fusion. This approach often suffers from loss of detail and insufficient accuracy in responding to objects of different scales. ASFFHead effectively overcomes these challenges by enhancing the aggregation capability of multi-level features and employing a selective fusion strategy. It improves the detection accuracy of large-scale objects and better captures the details of small-scale objects by utilizing feature maps at different levels, thereby enhancing the overall detection performance.

[0110] 4. Adaptive Medical Convolutional Module

[0111] This module integrates three branch structures: lightweight channel transform, depthwise separable convolution, and structure-aware convolution. Through multi-scale, multi-path information extraction mechanisms, it achieves accurate modeling and feature enhancement of irregular targets such as shoulder cysts. The specific structure is as follows: Figure 9 As shown.

[0112] In this module, the input feature map is simultaneously fed into three parallel branches to extract complementary features, with optimization designs for channel dimension compression and enhancement, local structure modeling, and boundary sensitivity modeling, respectively.

[0113] First, there's the lightweight channel transformation branch. This branch uses standard convolution operations to expand the channels of the input feature map, and further enhances its non-linear expressive power through batch normalization and activation functions. This branch is mainly used to capture importance information in the channel dimension, strengthen the network's modeling of high-value semantic features, and reduce the interference of redundant channels on model learning. The formula for expressing the above process is as follows:

[0114] F1=SiLU(BN(W1*X))(8)

[0115] Where F1 represents the feature map obtained after lightweight channel transformation; W1 is a 1×1 convolution kernel; BN represents the normalization operation; SiLU is the activation function; and the final output feature map has a size of W×H×C.

[0116] The second branch employs depthwise separable convolution with padding and stride values ​​of 1 for more refined feature mining, thereby capturing mid-scale cyst features. The expression after channel-wise convolution following depthwise separable convolution is:

[0117]

[0118] Among them, D c (i,j) represents the value of the c-th channel convolution output at position (i,j); K c (m,n) represents the kernel weight of the c-th channel, and (m,n) is the index of the position in the kernel; X c (i+m,j+n) represents the input features. Figure X c The pixel value at position (i+m,j+n); i and j represent the spatial coordinate indices in the current output feature map.

[0119] The weights K for all positions within the 3×3 convolution window c With input X c Perform pixel-by-pixel multiplication and summation to obtain the feature map D for each channel. c Then, the feature maps from each channel are fused through pixel-by-pixel convolution. The operation formula is as follows:

[0120] F2=SiLU(BN(W2*D))(10)

[0121] Where W2 represents the weight matrix used for linear transformation; D represents the set of feature maps formed by concatenating the convolution results of all channels.

[0122] The feature map F2 of the second branch is obtained, and its size is still W×H×C. The third branch introduces the structure-aware Snake convolution, which can adaptively adjust the sampling path based on the gradient information of the feature map and dynamically slide along the target edge to extract features. The Snake convolution has stronger adaptability and discriminative power when dealing with irregular edges, curved structures and atypical lesion regions. The formula for the above operation is expressed as Equation (11). After the convolution operation, normalization and activation function processing are performed as shown in Equation (12).

[0123]

[0124] F3=SiLU(BN(SnakeConv(X)))(12)

[0125] Where F3(i,j) represents the output value of the Snake convolution at position (i,j); K(m,n) represents the weight parameters of the convolution kernel at position (m,n); Ω represents the set of indices of the convolution kernel window; X(i+Δx) m,n ,j+Δy m,n ) represents the position at offset i+Δx in the input feature map. m,n ,j+Δy m,n Pixel value; Δx m,n Δy represents the horizontal offset corresponding to position (m,n) in the convolution kernel. m,n Let represent the vertical offset corresponding to position (m,n) in the convolution kernel; i, j represent the spatial coordinate indices in the current output feature map; F3 represents the complete feature map composed of the Snake convolution outputs at all positions; SnakeConv(X) represents the input feature map... Figure X Intermediate results obtained by performing Snake convolution operations;

[0126] After the three branches output, they are concatenated along the channel dimension to obtain a feature map with a scale of W×H×3C. Finally, a 1×1×2C convolution is used to complete information fusion and compression, and the final output is a feature map that fuses multi-dimensional information with a final output size of W×H×2C.

[0127] S3, Based on the segmentation results, calculate the defect area and defect ratio of the glenoid cavity;

[0128] Specifically, the core idea behind calculating the defect area and proportion is to predict the complete and actual morphology of the glenoid cavity using a deep learning model, calculate the pixel area difference between the two, and then derive the absolute area and relative percentage of the glenoid cavity defect. The normal glenoid cavity morphology is generated by the deep learning model based on training data. The glenoid cavity morphology in the patient's current state is obtained by segmenting the actual image using the deep learning model. The defect pixel area is the difference between the pixel area of ​​the complete glenoid cavity and the actual glenoid cavity. The defect proportion is the percentage of the defect area to the complete glenoid cavity area. To quantitatively assess the degree of glenoid cavity defect, this algorithm calculates the absolute area and relative percentage of the defect based on the complete glenoid cavity morphology predicted by the deep learning model and the actual glenoid cavity morphology. This calculation process combines deep learning segmentation results with mathematical geometric analysis, accurately quantifying the degree of damage and providing data support for subsequent clinical diagnosis and treatment.

[0129] Using a deep learning model and training data from healthy samples, the standard complete shape of the glenoid cavity is generated, denoted as M. full ,in:

[0130] M full (x,y)∈{0,1}

[0131] Among them, M full (x,y) = 1 indicates that pixel (x,y) belongs to the complete bone region.

[0132] The actual morphology of the glenoid cavity or humeral head extracted from the patient's current images is denoted as M. actual ,in:

[0133] M actual (x,y)∈{0,1}

[0134] Among them, M actual (x,y) = 1 indicates that pixel (x,y) belongs to the complete glenoid region.

[0135] The pixel area of ​​the defective region can be calculated using the pixel difference between the complete glenoid morphology and the actual glenoid morphology:

[0136]

[0137] The defect ratio is the percentage of the defected area to the intact glenoid cavity area, expressed as:

[0138]

[0139] S4. Based on the stated defect ratio, determine the degree of humeral head defect;

[0140] Specifically, based on the calculated defect area, this embodiment defines three levels of defect severity: minor defect, moderate defect, and severe defect. To determine the thresholds for these levels, we classify humeral head defects according to experience and medical standards.

[0141] Minor defect (Grade 1):

[0142] The proportion of the damaged area to the intact area is less than 10%, that is, P defect If the value is less than 0.1, it indicates a minor defect.

[0143] Moderate defect (Grade 2):

[0144] The proportion of the defective area to the intact area is between 10% and 30%, that is, 0.1 ≤ P defect If the value is less than 0.3, it is considered a moderate defect.

[0145] Severe defects (Grade 3):

[0146] The proportion of the damaged area to the intact area is greater than 30%, that is, P defect A value ≥0.3 indicates a severe defect.

[0147] Using the above calculation methods, doctors can accurately understand the extent of a patient's injury, thus providing a quantitative basis for treatment plans. Furthermore, this method has good universality and can be extended to other similar medical image segmentation and quantitative analysis tasks.

[0148] The effectiveness and superiority of the method of the present invention will be verified through experiments below.

[0149] 1. Introduction to the Data Experiment Set and Experimental Environment

[0150] In this embodiment, the experimental dataset, provided by a collaborating hospital, contains 800 MRI images of the glenoid fossa of the shoulder and 792 MRI images of the humeral head of the shoulder. These images were acquired using an ultrasound machine with an appropriate probe frequency, ensuring image quality and applicability. The dataset was labeled using Labelme image annotation software by two experienced sports medicine surgeons and rigorously reviewed by a kinesiology injury specialist to ensure accuracy. Labeling examples are provided. Figure 6 As shown.

[0151] To effectively train and evaluate the deep learning model, the experimental dataset was divided into training, validation, and test sets in an 8:1:1 ratio. 640 MRI images of the glenoid cavity and 633 MRI images of the humeral head were used as the training set for model training and optimization, while the remaining images were used in a 1:1 ratio for the validation and test sets.

[0152] 2. Evaluation Indicators

[0153] To evaluate the performance of this algorithm in defect area diagnosis, this embodiment employs several standard segmentation evaluation metrics to demonstrate the network's performance. In this work, this embodiment selects four metrics to evaluate performance: intersection over union (IoU), Dice similarity coefficient (DlCE), precision (PRE), Hausdorff distance (HD), and number of parameters. These metrics not only help measure the model's performance in segmentation accuracy but also provide in-depth analysis of segmentation boundaries, computational efficiency, and other aspects. These metrics are described below:

[0154]

[0155] HD = max{dis(P,G),dis(G,P)}

[0156] Where A represents the predicted region, B represents the true region, TP represents the true positive, FP represents the false positive, dis(P,G) represents the farthest distance from the predicted boundary point set P to the true boundary point set G, and dis(G,P) represents the farthest distance from the true boundary point set G to the predicted boundary point set P.

[0157] Intersection over Union (IoU) and Dice similarity coefficient are the most commonly used evaluation metrics in image segmentation tasks, widely applied in medical image segmentation. IoU, by calculating the intersection and union ratios of the predicted and ground truth regions, directly reflects the model's accuracy in target region overlap. Dice coefficient is another metric for evaluating region overlap, particularly suitable for tasks focusing on small target regions, and frequently used in medical imaging to evaluate the segmentation performance of organs or lesions. In this embodiment, these two metrics are used to evaluate the model's segmentation accuracy in the humeral head defect region. Furthermore, considering the complex shape and boundaries of the humeral head defect image in this embodiment, Hausdorff distance (HD) is used to evaluate the maximum distance between the predicted and ground truth boundaries. This metric is particularly effective in handling small-scale, complex segmentation tasks, helping us understand the model's performance in boundary accuracy. Precision, a classic evaluation metric, is also used in this embodiment to measure the model's segmentation accuracy and coverage. Precision reflects the proportion of regions predicted as positive by the model to actually be positive. These metrics help us balance false positives and false negatives, which is crucial in target region segmentation tasks.

[0158] 3. Comparative Experiment

[0159] In this embodiment, to verify the effectiveness of the proposed humeral head defect area diagnosis algorithm based on the feature enhancement module, we conducted a comparative experiment with several current mainstream medical image segmentation models. These models include: (1) U-Net, as a classic model in the field of medical image segmentation, has shown good performance in many medical imaging tasks. It extracts and reconstructs the details of the target region through an encoder-decoder structure, which is suitable for relatively simple segmentation tasks. We compare U-Net with the model proposed in this study to evaluate its accuracy and detail capture ability in processing glenoid defect images. (2) Mask R-CNN is an instance segmentation model based on Region Proposal Network (RPN), which has strong segmentation accuracy and boundary detection ability. Since humeral head defect images often contain complex boundaries and small targets, Mask R-CNN will be used as one of the comparison models to verify whether our proposed model is more advantageous in handling such tasks. (3) DeepLabV3+ adopts techniques such as dilated convolutions and spatial pyramid pooling (ASPP), which can effectively capture contextual information at different scales. This model has outstanding segmentation performance in complex scenes and is suitable for multi-scale feature extraction tasks. Therefore, it is an important benchmark model in our comparative experiments. (4) By comparing with these models, this embodiment will analyze in depth the performance differences of each model in the glenoid defect segmentation task, especially in terms of segmentation accuracy, boundary capture, and computational efficiency. (5) YOLOv11 is a newer version in the YOLO series, which further improves the target detection accuracy and inference speed. This version optimizes the target localization and multi-class object detection capabilities by integrating the latest neural network architecture and training techniques. Since humeral head defect images often contain complex boundaries and small lesion areas, YOLOv11 will be used as one of the comparative models to verify the relative performance of our proposed model in handling these complex scenarios. The experimental results will provide strong support for the advantages of the model proposed in this study and verify its feasibility and practicality in clinical applications. Five evaluation indicators, namely, coefficient, accuracy, HD distance, and number of running parameters, are used to quantitatively analyze the model.

[0160] Table 1. Comparative Experimental Results of Different Algorithms on Shoulder Defect Dataset

[0161]

[0162]

[0163] The experimental results in Table 1 demonstrate that the proposed model outperforms other state-of-the-art models in almost all evaluation metrics. In lesion image segmentation, this model achieves the best IoU (85.51%), accuracy (92.41%), and Hausdorff distance (1.077 mm). This achievement is primarily attributed to the introduction of a hierarchical cascaded feature fusion module in the upsampling part, which effectively aggregates multi-scale information, thereby improving segmentation accuracy. Simultaneously, the improved dynamic mask module helps the model focus on key information, showing significant advantages, especially in the segmentation of the sacroiliac joint with complex backgrounds. Furthermore, despite the significant improvement in segmentation performance, the number of model parameters is effectively controlled, even lower than other similar methods. This advantage is attributed to the introduction of partial convolution, which significantly reduces the number of floating-point operations and memory accesses, thereby improving computational efficiency.

[0164] 4. Ablation test

[0165] After introducing the EG-Module, the model's IoU improved to 87.54%, DICE slightly decreased to 93.77%, but PRE increased to 96.34%, and HD decreased to 2.4mm, indicating that this module helps enhance the ability to characterize local details in defect areas and improve prediction accuracy. Further adding ASFFHead increased IoU to 88.42%, DICE rose to 93.14%, PRE significantly increased to 97.29%, and HD decreased to 2.2mm, the lowest value to date, indicating that this module can effectively fuse feature information at different scales, especially performing well in edge region segmentation. When all three modules are integrated to form DDDNet, IoU reaches 89.77%, DICE improves to 94.71%, PRE is 98.16%, and HD remains at 2.3mm. The overall performance is better than any individual combination, fully validating the complementarity and synergistic enhancement effect of each module, providing strong support for refined medical image segmentation.

[0166] Table 2 Ablation Experiment Results

[0167]

[0168]

[0169] 5. Qualitative Results

[0170] Effects of glenoid segmentation Figure 7 As shown, the first column is the MRI image of the glenoid cavity, the second column is the image annotated by experts, and the rest are the segmentation results of various segmentation networks. Compared with other algorithms, the segmentation results of the algorithm in this paper have finer contours and are closer to the images annotated by experts.

[0171] The effect of segmentation of the humeral head is as follows Figure 8As shown, the first column contains MRI images of the humeral head, the second column contains images annotated by experts, and the remaining columns contain segmentation results from various segmentation networks.

[0172] Second Embodiment

[0173] This embodiment provides a diagnostic system for the area of ​​defects in the glenoid cavity and humeral head. The diagnostic system for the area of ​​defects in the glenoid cavity and humeral head includes the following modules:

[0174] The data acquisition module is used to acquire medical imaging data of the glenoid cavity and humeral head to be processed;

[0175] The target segmentation module is used to input the medical image data into a preset segmentation model to obtain the segmentation results of the glenoid cavity and the humeral head;

[0176] The data calculation module is used to calculate the defect area and defect ratio of the glenoid cavity based on the segmentation results;

[0177] The defect severity assessment module is used to determine the degree of humeral head defect based on the defect ratio.

[0178] It should be noted that the diagnostic system for the area of ​​glenoid and humeral head defects in this embodiment corresponds to the diagnostic method for the area of ​​glenoid and humeral head defects in the first embodiment described above. The functions implemented by each functional module in the diagnostic system for the area of ​​glenoid and humeral head defects in this embodiment correspond one-to-one with the process steps in the diagnostic method for the area of ​​glenoid and humeral head defects in the first embodiment described above; therefore, they will not be described again here.

[0179] Third Embodiment

[0180] This embodiment provides an electronic device, such as... Figure 10 As shown, the electronic device includes a processor and a memory; wherein the processor and the memory can be connected via a communication bus; the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment described above. Furthermore, the electronic device may also include a transceiver, the processor and the transceiver can be connected via a communication bus, and the transceiver is used to communicate with other devices.

[0181] Below, in conjunction with Figure 10 A detailed introduction to each component of this electronic device is provided below:

[0182] The processor is the control center of the electronic device. The electronic device may include multiple processors, each of which can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). The term "processor" can refer to a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), other general-purpose processors, application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), one or more field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor can perform various functions of the electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0183] In a specific implementation, as one example, the processor may include one or more CPUs, for example... Figure 10 CPU0 and CPU1 shown are, of course, merely illustrative examples.

[0184] The memory is used to store the software program that executes the solution of the present invention, and the processor controls its execution. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0185] Optionally, the memory may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory may be integrated with the processor or may exist independently, and may be accessed through the interface circuit of the electronic device (…). Figure 10 (Not shown in the image) is coupled to the processor; however, this embodiment of the invention does not impose specific limitations on this.

[0186] The transceiver may include a receiver and a transmitter. Figure 10 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function. The transceiver can be integrated with the processor or exist independently, and is connected through the interface circuit of the electronic device (…). Figure 10 (Not shown in the image) is coupled to the processor, and this embodiment of the invention does not specifically limit this.

[0187] In addition, it should be noted that, Figure 10 The structure of the electronic device shown is not intended to limit the device. Actual devices may include more or fewer components than shown, or combine certain components, or have different component arrangements. Furthermore, the technical effects achieved by this electronic device when performing the method of the first embodiment described above can be referenced to the technical effects described in the first embodiment; therefore, they will not be repeated here.

[0188] Fourth embodiment

[0189] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. The instruction stored therein can be loaded and executed by a processor in a terminal.

[0190] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely or partially hardware embodiment, a completely or partially software embodiment, or an embodiment combining software and hardware aspects. Moreover, when implemented in software, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any usable medium accessible to a computer or a data storage device such as a server or data center containing one or more sets of usable media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive (SSD).

[0191] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0192] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element. Furthermore, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Additionally, the character " / " in this text generally indicates an "or" relationship between the preceding and following objects, but it can also indicate an "AND / OR" relationship. Please refer to the context for specific interpretations. "At least one" refers to one or more items, while "more than" refers to two or more items. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can be represented as: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0194] Furthermore, it is understood that in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0195] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0196] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of functional modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, i.e., they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Additionally, the functional units in the various embodiments of this invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0197] If the method is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0198] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments of the present invention have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make several improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A diagnostic method for the area of ​​defects in the glenoid cavity and humeral head, characterized in that, include: Acquire medical imaging data of the glenoid cavity and humeral head to be processed; The medical image data is input into a preset segmentation model to obtain the segmentation results of the glenoid cavity and the humeral head; Based on the segmentation results, the defect area and defect ratio of the glenoid cavity are calculated; Based on the aforementioned defect ratio, the degree of humeral head defect is determined; The segmentation model includes multiple feature enhancement modules that constitute the Transformer structure; Starting from the second feature enhancement module, the output of each feature enhancement module is processed sequentially by the adaptive medical convolution module and the edge attention mechanism module, and the resulting feature data is input into the segmentation head. The segmentation head segments the glenoid cavity and humeral head based on multiple sets of input feature data and outputs the segmentation results of the glenoid cavity and humeral head. The feature enhancement module includes an overlapping block embedding module and multiple feature enhancement units; wherein, the multiple feature enhancement units are connected in series, and adjacent feature enhancement units are connected through a downsampling layer; The overlapping block embedding module introduces overlapping patch coding, and performs convolution on the input image with a kernel size of [missing value]. The convolutional layer transforms the network so that each vector in the result represents an embedded overlapping region with context, serving as a token sequence input to the Transformer structure; the input sequence is then fed into the feature enhancement unit. First, the output is obtained through a feedforward feedback network. The formula is expressed as: ； in, For learnable scaling parameters; It serves as a local convolutional feedback channel, used to embed spatial structure information; This indicates that the input is normalized; This indicates that each token is transformed independently, and the transformation operation is represented as follows: ； in, , Represents the weight matrix. This represents a non-linear activation function used to normalize the input; and Represents the bias vector; The system then enters the dimensional interaction module to enhance information exchange between tokens in different locations; subsequently, the cascaded group attention mechanism module processes the tokens from the dimensional interaction module. Divided into channel dimensions The system is divided into groups, and local self-attention is performed independently for each group, thus reducing computational cost. Finally, the outputs of each group are concatenated and then subjected to information fusion operation, which is used as the output of the feature enhancement unit. The data processing procedure of the edge attention mechanism module includes: The Sobel operator is used to extract the first-order gradient information of the input image in the horizontal and vertical directions, respectively; where the horizontal operator... and vertical direction operator Represented as: ； Each channel of the input image is respectively compared with Perform a convolution operation to obtain the horizontal gradient map, and then convolve each channel of the input image with... Perform a convolution operation to obtain the gradient map in the vertical direction; The horizontal and vertical gradient maps are processed separately using a global average pooling layer to obtain the aggregate tensor of the input image in the vertical direction. and the aggregation tensor in the horizontal direction ; Will and Multiplying them together yields a location-dependent attention map M; Enter M Convolutional layer, then The output of the convolutional layer is added to the input image of the edge attention mechanism module to obtain the optimized feature map; The data processing procedure for the segmentation head includes: Multiple sets of input feature data are fused to obtain fused feature data. ; The weight of each pixel is dynamically calculated; the weight generation formula is as follows: ； in, For dynamically generated weight masks; This represents a non-linear activation function used to normalize the input; Indicates the convolution operation; The range is [0,1], and a larger value indicates that the model pays more attention to the corresponding region; according to The final dynamic mask is generated to guide segmentation; the formula for generating the dynamic mask is: ； in, For dynamic masks; These are the preliminary predictions from the model; This indicates pixel-wise weighting; Based on the generated dynamic mask, segmentation of the glenoid cavity and humeral head is achieved.

2. The diagnostic method for defects in the glenoid cavity and humeral head as described in claim 1, characterized in that, The number of feature enhancement units is three; in the first feature enhancement unit Set to 16, in the second feature enhancement unit Set to 8, in the last feature enhancement unit Set it to 4.

3. The diagnostic method for the area of ​​defects in the glenoid cavity and humeral head as described in claim 1, characterized in that, The operational formula for local self-attention is expressed as follows: ； in, Indicates the first A query matrix; For the first A key matrix; For the first A Value matrix; The dimension of the Key; T Represents the transpose of a matrix; Indicates the first The output of each attention head; This indicates normalization processing.

4. The diagnostic method for defects in the glenoid cavity and humeral head as described in claim 1, characterized in that, Based on the segmentation results, the defect area and defect ratio of the glenoid cavity are calculated, including: Using the segmentation model, a standard complete shape of the glenoid cavity is generated based on the training data of healthy samples; Using the segmentation model, the actual shape of the glenoid cavity is generated based on the medical imaging data; The pixel area difference between the standard complete shape and the actual shape is calculated to obtain the defect area of ​​the glenoid cavity; Calculate the percentage of the defect area to the intact glenoid cavity area to obtain the defect ratio of the glenoid cavity.

5. The diagnostic method for the area of ​​defects in the glenoid cavity and humeral head as described in claim 1, characterized in that, Based on the aforementioned defect ratio, the degree of humeral head defect is determined, including: If the defect rate is less than 10%, it is determined to be a minor defect; If the defect rate is between 10% and 30%, it is determined to be a moderate defect; If the defect rate is greater than 30%, it is determined to be a severe defect.

6. A diagnostic system for defects in the glenoid cavity and humeral head, characterized in that, include: The data acquisition module is used to acquire medical imaging data of the glenoid cavity and humeral head to be processed; The target segmentation module is used to input the medical image data into a preset segmentation model to obtain the segmentation results of the glenoid cavity and the humeral head; The data calculation module is used to calculate the defect area and defect ratio of the glenoid cavity based on the segmentation results; The defect severity assessment module is used to determine the degree of humeral head defect based on the defect ratio. The segmentation model includes multiple feature enhancement modules that constitute a Transformer structure; starting from the second feature enhancement module, the output of each feature enhancement module is processed sequentially by an adaptive medical convolution module and an edge attention mechanism module, and the resulting feature data is input into the segmentation head. The segmentation head segments the glenoid cavity and humeral head based on multiple sets of input feature data and outputs the segmentation results of the glenoid cavity and humeral head. The feature enhancement module includes an overlapping block embedding module and multiple feature enhancement units; wherein, the multiple feature enhancement units are connected in series, and adjacent feature enhancement units are connected through a downsampling layer; The overlapping block embedding module introduces overlapping patch coding, and performs convolution on the input image with a kernel size of [missing value]. The convolutional layer transforms the network so that each vector in the result represents an embedded overlapping region with context, serving as a token sequence input to the Transformer structure; the input sequence is then fed into the feature enhancement unit. First, the output is obtained through a feedforward feedback network. The formula is expressed as: ； in, For learnable scaling parameters; It serves as a local convolutional feedback channel, used to embed spatial structure information; This indicates that the input is normalized; This indicates that each token is transformed independently, and the transformation operation is represented as follows: ； in, , Represents the weight matrix. This represents a non-linear activation function used to normalize the input; and Represents the bias vector; The system then enters the dimensional interaction module to enhance information exchange between tokens in different locations; subsequently, the cascaded group attention mechanism module processes the tokens from the dimensional interaction module. Divided into channel dimensions The system is divided into groups, and local self-attention is performed independently for each group, thus reducing computational cost. Finally, the outputs of each group are concatenated and then subjected to information fusion operation, which is used as the output of the feature enhancement unit. The data processing procedure of the edge attention mechanism module includes: The Sobel operator is used to extract the first-order gradient information of the input image in the horizontal and vertical directions, respectively; where the horizontal operator... and vertical direction operator Represented as: ； Each channel of the input image is respectively compared with Perform a convolution operation to obtain the horizontal gradient map, and then convolve each channel of the input image with... Perform a convolution operation to obtain the gradient map in the vertical direction; The horizontal and vertical gradient maps are processed separately using a global average pooling layer to obtain the aggregate tensor of the input image in the vertical direction. and the aggregation tensor in the horizontal direction ; Will and Multiplying them together yields a location-dependent attention map M; Enter M Convolutional layer, then The output of the convolutional layer is added to the input image of the edge attention mechanism module to obtain the optimized feature map; The data processing procedure for the segmentation head includes: Multiple sets of input feature data are fused to obtain fused feature data. ; The weight of each pixel is dynamically calculated; the weight generation formula is as follows: ； in, For dynamically generated weight masks; This represents a non-linear activation function used to normalize the input; Indicates the convolution operation; The range is [0,1], and a larger value indicates that the model pays more attention to the corresponding region; according to The final dynamic mask is generated to guide segmentation; the formula for generating the dynamic mask is: ； in, For dynamic masks; These are the preliminary predictions from the model; This indicates pixel-wise weighting; Based on the generated dynamic mask, segmentation of the glenoid cavity and humeral head is achieved.

Images