CN-120809145-B - Diagnostic method and system for defect area of glenoid and humeral head

CN120809145BCN 120809145 BCN120809145 BCN 120809145BCN-120809145-B

Abstract

The invention discloses a diagnosis method and a diagnosis system for the defect area of a glenoid and a humeral head, which belong to the technical field of image processing, wherein the method comprises the steps of obtaining medical image data of the glenoid and the humeral head to be processed; the method comprises the steps of inputting medical image data into a preset segmentation model to obtain a segmentation result of the glenoid and the humerus, calculating the defect area and the defect proportion of the glenoid based on the segmentation result, and determining the defect degree of the humerus based on the defect proportion. The diagnosis method and the diagnosis system for the defect areas of the glenoid and the humeral head provide an efficient, accurate and intelligent diagnosis auxiliary tool for clinic, and have good application prospect and popularization value.

Inventors

GUO JIN
HOU HAONAN
LIU ZHENLONG
FANG JINGCHAO
LI MENGQI
SUN SHUYING

Assignees

北京科技大学
北京大学第三医院（北京大学第三临床医学院）

Dates

Publication Date: 20260508
Application Date: 20250625

Claims (6)

1. A method of diagnosing a defect area of a glenoid and humeral head, comprising: acquiring medical image data of a glenoid and a humeral head to be processed; Inputting the medical image data into a preset segmentation model to obtain a segmentation result of the glenoid and the humeral head; calculating a defect area and a defect proportion of the glenoid based on the segmentation result; determining a degree of defect of the humeral head based on the defect ratio; The segmentation model comprises a plurality of feature enhancement modules forming a transducer structure, wherein from a second feature enhancement module, the output of each feature enhancement module is respectively processed by a self-adaptive medical convolution module and an edge attention mechanism module in sequence, and the obtained feature data is input into a segmentation head; The characteristic enhancement module comprises an overlapped block embedding module and a plurality of characteristic enhancement units, wherein the plurality of characteristic enhancement units are connected in series, and two adjacent characteristic enhancement units are connected through a downsampling layer; The overlapped block embedding module introduces overlapped patch coding to carry out convolution kernel size on the input image to be The convolution layer transform, each vector in the result representing a Token sequence with overlapping region embedding of context as Transfomer structure input, input sequence to feature enhancement unit Firstly, obtaining output through a feedforward feedback network The formula is: ; Wherein, the Is a learnable scaling parameter; is a local convolution feedback channel and is used for embedding spatial structure information; representation normalizes the input; Representation of transforming each Token independently, its transformation operation is represented as: ; Wherein, the , The weight matrix is represented by a matrix of weights, Representing a nonlinear activation function for normalizing the input; And Representing the bias vector; then entering a secondary element interaction module to enhance the direct information exchange of Token between different positions, and then processing the secondary element interaction module by a cascade grouping attention mechanism module Division into channel dimensions Grouping, then independently executing local self-attention on each group, and reducing the calculated amount; finally, carrying out information fusion operation after the outputs of the groups are connected in series to serve as the output of the characteristic enhancement unit; the data processing process of the edge attention mechanism module comprises the following steps: Extracting one-step degree information of an input image in a horizontal direction and a vertical direction respectively by using a Sobel operator, wherein the operator in the horizontal direction And vertical direction operator Expressed as: ; associating each channel of the input image with a respective one of the channels Performing convolution operation to obtain horizontal gradient image, and respectively combining each channel of input image with each other Performing convolution operation to obtain a gradient map in the vertical direction; The global average pooling layer is used for respectively processing the horizontal gradient map and the vertical gradient map to obtain the aggregation tensor of the input image in the vertical direction And polymerization tensor in horizontal direction ; Will be And Multiplying to obtain a position-dependent attention map M; Inputting M into Convolving layer and then The output of the convolution layer is added with the input image of the edge attention mechanism module to obtain an optimized feature map; the data processing process of the dividing head comprises the following steps: fusing the input multiple groups of characteristic data to obtain fused characteristic data ; Dynamically calculating the weight of each pixel, wherein the weight generation formula is as follows: ; Wherein, the A weight mask dynamically generated; Representing a nonlinear activation function for normalizing the input; representing a convolution operation; the larger the value indicates that the model focuses more on the corresponding region; According to Generating a final dynamic mask for guiding segmentation, wherein the dynamic mask generation formula is as follows: ; Wherein, the Is a dynamic mask; Preliminary prediction results of the model; Representing pixel-by-pixel weighting; based on the generated dynamic mask, a segmentation of the glenoid and humeral head is achieved.
2. The method for diagnosing a defective area of a glenoid and humeral head as claimed in claim 1, wherein the number of the feature-enhancing units is 3, in the first feature-enhancing unit Set to 16, in a second feature enhancement unit Set to 8, in the last feature enhancement unit Set to 4.
3. The diagnostic method for a glenoid and humeral head defect area of claim 1, wherein the local self-attention operational formula represents: ; Wherein, the Represent the first A plurality of Query matrices; Is the first A plurality of Key matrices; Is the first A Value matrix; t represents the transpose of the matrix; Represent the first Output results of the individual attention heads; the normalization process is represented.
4. The method for diagnosing a defective area of a glenoid and a humeral head according to claim 1, wherein calculating a defective area and a defective proportion of the glenoid based on the segmentation result includes: generating a standard complete form of the glenoid from training data of the health sample using the segmentation model; generating an actual form of the glenoid from the medical image data using the segmentation model; calculating the pixel area difference between the standard complete form and the actual form to obtain the defect area of the glenoid; and calculating the percentage of the defect area to the whole glenoid area to obtain the defect proportion of the glenoid.
5. The diagnostic method for a glenoid and humeral head defect area of claim 1, wherein determining the degree of humeral head defect based on the defect ratio includes: if the defect proportion is less than 10%, judging that the defect is slight; If the defect ratio is between 10% and 30%, then determining a moderate defect; If the defect ratio is greater than 30%, it is determined that the weight is defective.
6. A diagnostic system for a glenoid and humeral head defect area, comprising: the data acquisition module is used for acquiring medical image data of the joint spittoon and the humerus head to be processed; The target segmentation module is used for inputting the medical image data into a preset segmentation model to obtain a segmentation result of the glenoid and the humerus head; a data calculation module for calculating a defect area and a defect proportion of the glenoid based on the segmentation result; A defect degree judging module for determining the defect degree of the humeral head based on the defect proportion; The segmentation model comprises a plurality of feature enhancement modules forming a transducer structure, wherein from a second feature enhancement module, the output of each feature enhancement module is respectively processed by a self-adaptive medical convolution module and an edge attention mechanism module in sequence, and the obtained feature data is input into a segmentation head; The characteristic enhancement module comprises an overlapped block embedding module and a plurality of characteristic enhancement units, wherein the plurality of characteristic enhancement units are connected in series, and two adjacent characteristic enhancement units are connected through a downsampling layer; The overlapped block embedding module introduces overlapped patch coding to carry out convolution kernel size on the input image to be The convolution layer transform, each vector in the result representing a Token sequence with overlapping region embedding of context as Transfomer structure input, input sequence to feature enhancement unit Firstly, obtaining output through a feedforward feedback network The formula is: ; Wherein, the Is a learnable scaling parameter; is a local convolution feedback channel and is used for embedding spatial structure information; representation normalizes the input; Representation of transforming each Token independently, its transformation operation is represented as: ; Wherein, the , The weight matrix is represented by a matrix of weights, Representing a nonlinear activation function for normalizing the input; And Representing the bias vector; then entering a secondary element interaction module to enhance the direct information exchange of Token between different positions, and then processing the secondary element interaction module by a cascade grouping attention mechanism module Division into channel dimensions Grouping, then independently executing local self-attention on each group, and reducing the calculated amount; finally, carrying out information fusion operation after the outputs of the groups are connected in series to serve as the output of the characteristic enhancement unit; the data processing process of the edge attention mechanism module comprises the following steps: Extracting one-step degree information of an input image in a horizontal direction and a vertical direction respectively by using a Sobel operator, wherein the operator in the horizontal direction And vertical direction operator Expressed as: ; associating each channel of the input image with a respective one of the channels Performing convolution operation to obtain horizontal gradient image, and respectively combining each channel of input image with each other Performing convolution operation to obtain a gradient map in the vertical direction; The global average pooling layer is used for respectively processing the horizontal gradient map and the vertical gradient map to obtain the aggregation tensor of the input image in the vertical direction And polymerization tensor in horizontal direction ; Will be And Multiplying to obtain a position-dependent attention map M; Inputting M into Convolving layer and then The output of the convolution layer is added with the input image of the edge attention mechanism module to obtain an optimized feature map; the data processing process of the dividing head comprises the following steps: fusing the input multiple groups of characteristic data to obtain fused characteristic data ; Dynamically calculating the weight of each pixel, wherein the weight generation formula is as follows: ; Wherein, the A weight mask dynamically generated; Representing a nonlinear activation function for normalizing the input; representing a convolution operation; the larger the value indicates that the model focuses more on the corresponding region; According to Generating a final dynamic mask for guiding segmentation, wherein the dynamic mask generation formula is as follows: ; Wherein, the Is a dynamic mask; Preliminary prediction results of the model; Representing pixel-by-pixel weighting; based on the generated dynamic mask, a segmentation of the glenoid and humeral head is achieved.

Description

Diagnostic method and system for defect area of glenoid and humeral head Technical Field The invention relates to the technical field of image processing, in particular to a diagnosis method and a diagnosis system for the defect area of a glenoid and a humeral head. Background Dislocation of the shoulder is a common traumatic disease in the clinic, often resulting in bone damage to the humeral head and glenoid. Humeral head defect and glenoid defect are two common pathologies of shoulder joint diseases, seriously affect the stability and function of shoulder joints of patients, and bring great trouble to the daily life and exercise capacity of patients. With increasing exercise intensity and traumatic incidents, the incidence of humeral head and glenoid defects has increased year by year. Humeral head defects are primarily manifested as local bone injuries or deletions, whereas glenoid defects are manifested as partial or complete glenoid deletions, which may cause a series of subsequent problems of joint degeneration, arthritis, etc. Thus, early detection and accurate assessment of the area of these defects can provide an important basis for clinical treatment and intervention. Traditional humeral head and glenoid defect diagnosis methods rely on X-ray, CT scan and MRI images to determine the extent and extent of injury by manual analysis by a specialist. However, these methods are susceptible to interference from physician experience, image noise, and complex morphology, resulting in poor diagnostic accuracy and efficiency, and particularly when processing large-scale images, the physician's judgment may deviate, affecting the treatment decision. Disclosure of Invention The invention provides a diagnosis method and a diagnosis system for the defect area of a glenoid and a humeral head, which aim to solve the problem that the traditional diagnosis method for the defects of the humeral head and the glenoid relies on manual analysis to judge the scope and degree of damage. Thus, the diagnosis accuracy and efficiency are low, and the technical problem of treatment decision is affected. In order to solve the technical problems, the invention provides the following technical scheme: in one aspect, the present invention provides a method of diagnosing a glenoid and humeral head defect area, the method comprising: acquiring medical image data of a glenoid and a humeral head to be processed; Inputting the medical image data into a preset segmentation model to obtain a segmentation result of the glenoid and the humeral head; calculating a defect area and a defect proportion of the glenoid based on the segmentation result; based on the defect ratio, a degree of the defect of the humeral head is determined. The segmentation model further comprises a plurality of feature enhancement modules forming a transducer structure, from the second feature enhancement module, the output of each feature enhancement module is processed by the self-adaptive medical convolution module and the edge attention mechanism module respectively, the obtained feature data are input into the segmentation head, the segmentation head realizes the segmentation of the glenoid and the humeral head based on the input multiple groups of feature data, and the segmentation result of the glenoid and the humeral head is output. The feature enhancement module comprises an overlapped block embedding module and a plurality of feature enhancement units, wherein the feature enhancement units are connected in series, and two adjacent feature enhancement units are connected through a downsampling layer; The overlapped block embedding module introduces overlapped patch coding, carries out convolution kernel size 7×7 convolution layer transformation on an input image, embeds each vector in the result representing an overlapped area with a context as a Token sequence input by Transfomer structures, inputs the Token sequence x to a characteristic enhancement unit, firstly obtains an output x' through a feedforward feedback network, and the formula is expressed as: x′=x+γ·Conv1×1(LN(x))+FFN(x) Wherein, gamma is a leachable scaling parameter, conv 1×1 is a local convolution feedback channel used for embedding space structure information, LN (x) represents normalizing input, FFN represents independently transforming each Token, and the transformation operation is expressed as follows: FFN(x)=W2·σ(W1x+b1)+b2 Wherein W 1,W2 represents a weight matrix, σ represents a nonlinear activation function for normalizing the input, b 1 and b 2 represent bias vectors; the method comprises the steps of firstly, carrying out a hierarchical grouping attention mechanism module, wherein x 'enters a secondary element interaction module to enhance Token direct information exchange among different positions, then dividing the x' processed by the secondary element interaction module into G groups in a channel dimension, and then independently executing local self-attention on each group a