A method and system for classifying developmental dysplasia of the hip

Abstract

The application discloses a multi-subgroup hip dysplasia classification method and system, and belongs to the field of artificial intelligence. The method comprises the following steps: collecting a pelvis orthographic X-ray film image of a patient, and pre-processing the image; constructing a coarse classification network, extracting general features in the image to obtain a coarse classification result of roughly identifying the category of hip dysplasia, wherein the general features comprise contour and texture; constructing a fine classification network, based on a graph convolution network, adding expression information of anatomical key points to the coarse classification result, and refining the classification result, wherein the anatomical key points comprise the highest point of the ilium, the outermost edge of the acetabulum, the center point of the femoral head, the tear drop point and the lowest point of the ischial tuberosity; and outputting the fine classification network result. The application can improve the classification efficiency and accuracy while improving the system generalization.

Description

Technical Field

[0001] This invention generally relates to the field of artificial intelligence technology. More specifically, this invention relates to a multi-subgroup hip dysplasia classification system. Background Technology

[0002] Developmental dysplasia of the hip (DDH) refers to various deformities that occur during hip joint development, including femoral head dysplasia, acetabular dysplasia, or hip dislocation. Furthermore, DDH is one of the leading causes of hip osteoarthritis. Among patients over 50 years of age with hip osteoarthritis, 25%–50% are due to DDH. Therefore, effective early diagnosis and intervention can not only alleviate DDH symptoms and improve work capacity, but also delay or prevent the progression of osteoarthritis, avoiding the need for total hip replacement surgery and effectively improving medical and economic efficiency.

[0003] Imaging studies are crucial for the early diagnosis and classification of diabetic dysplasia of the hip (DDH), as well as for severity analysis and treatment selection. Pelvic X-rays, particularly anteroposterior views, are the preferred auxiliary examination for DDH diagnosis and treatment, primarily used to analyze the weight-bearing area of ​​the acetabulum, femoral head dislocation, and the degree of joint degeneration. DDH classification differs between children and adults, employing different classification systems. For example, the Tönnis classification is commonly used for children, while the Crowe classification is more common for adults. While the definitions of these classification systems vary, they are all based on the relative position of the femoral head and acetabulum (pelvis), and both the Tönnis and Crowe classifications categorize DDH into types I-IV. To achieve accurate diagnosis of DDH, key indicators such as the center-edge angle (CE angle), the Tonnis angle, and the Sharp angle are measured clinically.

[0004] Currently, the diagnosis of DDH on anteroposterior pelvic X-rays is mostly done by doctors manually marking key anatomical points and measuring relevant parameters. This process is inefficient and highly dependent on the doctor's experience. For adult DDH patients with a longer disease duration, the lesions are often more severe, and junior doctors frequently encounter situations where they cannot determine the center of femoral head movement or locate the lateral edge of the true acetabulum due to severe femoral head deformation. In addition, because the femoral head of infants and adolescents is not fully developed and often exists in a transitional form of Y-shaped cartilage, there are wide individual differences among minors of different ages and developmental stages, resulting in low consistency in the selection of key points. These reasons directly lead to the current lack of a rapid, automated, and repeatable diagnostic method for DDH in clinical practice. Summary of the Invention

[0005] To address at least some of the problems in the background art regarding the large differences in DDH among different subgroups (ages), classification difficulties, and low generalization, this invention proposes a multi-subgroup hip dysplasia classification system based on graph convolutional networks. This system employs a multi-level classification framework, combining a point detection model and a graph convolutional network (GCN) to perform secondary classification on different subgroups and complex DDH data, improving classification efficiency and accuracy while enhancing system generalization. Therefore, this invention provides solutions in the following aspects. The first aspect of this invention provides a multi-subgroup hip dysplasia classification method, comprising: acquiring anteroposterior X-ray images of the patient's pelvis, and preprocessing the anteroposterior X-ray images to obtain preprocessing results to exclude the influence of non-hip dysplasia cases; constructing a coarse classification network, using a lightweight model, to extract common features from the images to obtain a coarse classification result that identifies the approximate category of hip dysplasia, the common features including contour and texture; constructing a fine classification network, based on a graph convolutional network, to refine the classification result by adding additional anatomical key point expression information to the coarse classification result, the anatomical key points including the highest point of the iliac bone, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity; and outputting the fine classification network results for analysis and use by doctors.

[0006] In one embodiment, the method further includes: obtaining the anatomical key points on the femoral head, acetabulum, and pelvis through a neural network based on the preprocessing results, wherein the center point of the femoral head in children is identified as the ossification nucleus of the femoral head, and if the ossification nucleus does not exist, the midpoint of the uppermost edge of the femur is used.

[0007] In one embodiment, the last layer of the coarse classification network uses the Softmax function to transform the output into a probability distribution. The maximum value in the output probability distribution is the confidence score of the classification result. Data with a confidence score lower than a set threshold are set to enter the subsequent fine classification stage for further refinement. Data with a confidence score higher than the set threshold are entered into the subsequent fine classification according to a set proportion to ensure the generalization of the fine classification model.

[0008] In one embodiment, the fine classification network defines the anatomical key points as graph nodes and the adjacency relationships between nodes as edges; it employs a graph convolutional network to learn the embedded representations of the nodes, including node feature information, adjacency relationships, and multi-hop propagation; it obtains the feature representations between key points, then fuses the image features of the encoding layers in the coarse classification network, and finally performs fine classification evaluation through a classifier.

[0009] In one embodiment, the coarse classification network employs lightweight models such as MobileNet and SqueezeNet to ensure faster processing speed.

[0010] In one embodiment, obtaining relevant anatomical points on the femoral head, acetabulum, and pelvis via neural networks includes using Faster R-CNN, UNet-like networks, or Transform-like networks.

[0011] In one embodiment, the fusion is performed using concatenation or weighted summation; the classifier includes a fully connected layer or a softmax layer.

[0012] In one embodiment, the image preprocessing includes excluding the following: hip dysplasia that has undergone surgical treatment, combined with other hip joint diseases, pelvic internal or external rotation at the time of imaging, and pelvic, iliac, or femoral incompleteness.

[0013] In one embodiment, the image preprocessing further includes denoising, window width / window level adjustment, and histogram equalization of the received image to optimize image quality.

[0014] A second aspect of the present invention provides a multi-subgroup hip dysplasia classification system, using the multi-subgroup hip dysplasia classification method described in any of the preceding claims.

[0015] This invention employs a multi-stage classification framework, combining a point detection model and a graph convolutional network (GCN), to perform secondary classification on different subgroups of people and complex DDH data. This improves classification efficiency and accuracy while enhancing system generalization. Compared to traditional indicator measurement methods, the advantages of this invention are: 1) It uses a deep learning classification network as the basic framework to achieve end-to-end output of DDH diagnosis; 2) It uses a multi-stage classification framework to solve the problem of difficult classification of complex data, improving classification efficiency while enhancing accuracy and reliability; 3) Addressing the challenge of significant differences in hip joint anatomy among different subgroups, it uses a graph convolutional network to automatically learn the representational connections of anatomical key points, effectively integrating the abstract features of the classification network with the association features of subgroups, improving the accuracy of personalized automatic diagnosis of DDH among different subgroups and enhancing clinical diagnostic efficiency. Attached Figure Description

[0016] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This illustrates a multi-subgroup hip dysplasia classification method according to an embodiment of the present invention; Figure 2 This illustrates a multi-subgroup hip dysplasia classification process according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating key points of the hip joint according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating the overall process of a graph-based integral network according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the edge connection according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be understood that the terms "first," "second," "third," and "fourth," etc., in the claims, specification, and drawings of this invention are used to distinguish different objects, rather than to describe a specific order. The terms "comprising" and "including" used in the specification and claims of this invention indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]." The first aspect of this invention provides a multi-subgroup hip dysplasia classification system method. The method employs a multi-level classification framework, combining a point detection model and a graph convolutional network (GCN), enabling secondary classification of different subgroup populations and complex DDH data. This improves classification efficiency and accuracy while enhancing the system's generalization ability.

[0018] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Figure 1 , 2 The diagram shown is a method diagram and a flowchart of the operation of the present invention.

[0019] From the appendix Figure 1 The multi-subgroup hip dysplasia classification method of the present invention specifically includes steps 100-400, including: Step 100: Acquire an anteroposterior X-ray image of the patient's pelvis, and obtain a preprocessing result by preprocessing the anteroposterior X-ray image of the pelvis to exclude the influence of non-hip dysplasia cases; Step 200: Construct a coarse classification network using a lightweight model to extract common features from the image to obtain a coarse classification result that roughly identifies the category of hip dysplasia. The common features include contour and texture. Step 300: Construct a fine classification network. Based on a graph convolutional network, the classification results are refined by adding additional anatomical key point expression information on the coarse classification results. The anatomical key points include the highest point of the ilium, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity. Step 400: Output the detailed classification network results for doctors to analyze and use.

[0020] Furthermore, by appendix Figure 2 This invention employs a multi-stage classification framework. The first stage is a coarse classification network, which outputs preliminary classification results and their confidence scores. This coarse classification quickly filters out approximate categories, reducing the computational burden of subsequent stages. Each classification result provides a confidence score to indicate its reliability; data with low confidence scores are more likely to be included in the second-stage classification training. The second stage primarily uses a fine classification network framework to handle complex data with low confidence scores from the coarse classification. The network input consists of two parts: anatomical keypoints obtained from the point detection model and image features from the coarse classification network. The graph convolutional network embeds the nodes and edges of the anatomical keypoints, then fuses the image features from the coarse classification model to refine the classification results, improving accuracy and reliability.

[0021] This invention first performs preliminary classification on anteroposterior X-ray images of the pelvis. For data with low confidence in the classification results, a secondary classification is performed. The main process includes: 1) Image acquisition and preprocessing M anteroposterior X-ray images of the pelvis were obtained from patients with different equipment from various manufacturers and of different ages. Exclusion criteria were: (1) hip dysplasia that had undergone surgical treatment; (2) other hip joint diseases; (3) pelvic internal or external rotation at the time of imaging; (4) pelvic, iliac, or femoral insufficiency. Preprocessing module: The received images were processed by denoising, window width / window level adjustment, histogram equalization, etc., to optimize image quality and facilitate subsequent network learning.

[0022] The training dataset is divided into a training set, a validation set, and a test set. The training set, validation set, and test set are independent of each other, and their sizes are m1, m2, and m3, respectively, which are natural numbers. Furthermore, m1 + m2 + m3 = M, and m1 ≥ 2 / 3 M. 2) Basic coarse classification framework The first-stage coarse classification network, M1, uses a lightweight model such as MobileNet or SqueezeNet to ensure fast processing speed. This stage mainly extracts common features from the image, such as contours and textures. These features help to quickly identify the approximate category (DDH / non-DDH, children / adults, simple data / complex data, etc.). The last layer typically uses the Softmax function to transform the output into a probability distribution. The maximum value in the output probability distribution is the confidence score of the classification result. Data with low confidence (e.g., below 0.5) (which are likely difficult to classify in DDH) will enter the subsequent fine classification stage for further refinement. Data with high confidence will also enter the subsequent fine classification stage to ensure the generalization of the fine classification model.

[0023] 3) Fine-classification framework based on graph convolutional networks The second-stage fine-classification framework is mainly based on graph convolutional networks. It refines the classification results by adding additional information about key points. The input data in this stage mainly consists of data with low confidence (e.g., below 0.5) from the coarse classification results. At the same time, data with higher confidence will also enter the next stage with a certain probability (e.g., 10%) to increase the generalization ability of the subsequent model.

[0024] a) Key Point Detection Module: The preprocessed image is processed through networks such as Faster R-CNN, UNet-like, or Transform-like to obtain relevant anatomical points on the femoral head, acetabulum, and pelvis, including: the highest point of the ilium, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity. Specific results can be seen in [link to documentation]. Figure 3 The diagram shows key points of the hip joint. In children, the center point of the femoral head is identified as the ossification nucleus of the femoral head. If the ossification nucleus does not exist, the midpoint of the uppermost edge of the femur is used.

[0025] b) Graph integral networks: such as Figure 4 The diagram shown is an overall flowchart of a graph convolutional integral network according to an embodiment of the present invention. (See attached diagram.) Figure 4 It can be seen that the input of the graph convolutional network M2 consists of two parts: 1) the grouped keypoints obtained by the midpoint detection network mentioned above (a), and 2) the image features of the first-stage coarse classification network M1. Among them, structural keypoints (the highest point of the iliac bone, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity, etc.) are defined as graph nodes, and the adjacency relationships between nodes are defined as edges. A graph convolutional network (GCN) is used to learn the embedded representations of nodes, including node feature information, adjacency relationships, multi-hop propagation, etc. Figure 4 The image shown is a schematic diagram of nodes and edge connections in graph convolution. It should be noted that this image is for illustrative purposes only, and the actual learning outcomes of the graph network may differ. Feature representations between key points are obtained, and then the image features from the coarse classification network encoding layer in step 2) above are fused (fusion methods include concatenation and weighted summation). Finally, a classifier (such as a fully connected layer or Softmax) is used for refined classification evaluation.

[0026] 4) Output of classification results The first stage, a coarse classification network, primarily extracts general features at the image level to obtain a rough classification result and its confidence level for DDH data. After completing the coarse classification, targeted feature learning and result analysis can be performed on data with low confidence levels. The second stage uses a graph convolutional network to distinguish the differences in anatomical key points across different data sets, thereby refining the classification results for data with low confidence levels (maintaining the results of the first stage or correcting the DDH subtyping). Finally, the network outputs refined classification results for doctors to analyze and use.

[0027] In a specific embodiment of the present invention, the multi-subgroup hip dysplasia classification method includes the following steps: 1) Generate training dataset This embodiment collected anteroposterior images of the hip joint from N (N>=5000) patients (including those using equipment from different manufacturers, of different genders, ages, and classifications). Inclusion criteria were as follows: 1. Complete pelvic and upper / middle third of the femur were included; 2. The sacroiliac joint was clearly and completely visualized; 3. Bilateral iliac wing and other bones showed uniform density, with clear trabeculae; 4. No pelvic rotation and lower limb internal rotation of 15°~20° were observed during imaging. Key points are marked as shown in the attached figure. Figure 1 The subtyping was done by physicians based on the patient's age, using the Tönnis subtyping (children) and the Crowe subtyping (adults). Then, all data underwent denoising, window width / level adjustment, and histogram equalization to obtain the original dataset for subsequent deep learning training (X1) and testing (X2).

[0028] 2) First stage coarse classification The first-stage coarse classification network, M1, uses a lightweight model such as MobileNet or SqueezeNet to ensure fast processing speed. This stage mainly extracts common features from the image, such as contours and textures. These features help to quickly identify the approximate category (DDH / non-DDH, children / adults, simple data / complex data, etc.). The final layer typically uses the Softmax function to transform the output into a probability distribution. The maximum value in the output probability distribution is the confidence score of the classification result. Data with lower confidence scores (most likely data that is difficult to classify in DDH) will receive greater weight in subsequent fine-grained classification stages for further refinement.

[0029] 3) The second-stage graph convolutional network fine classification framework The second-stage fine-classification framework is mainly based on graph convolutional networks. It refines the classification results by adding additional information about key points. The input to this stage mainly consists of data with low confidence from the coarse classification results. At the same time, data with high confidence will also enter the next stage with a certain probability (e.g., 10%) to increase the generalization ability of the subsequent model.

[0030] a. Key Point Detection Module: The preprocessed image is processed through networks such as Faster R-CNN, UNet-like, or Transform-like to obtain relevant anatomical points on the femoral head, acetabulum, and pelvis, including: the highest point of the ilium, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity (see details). Figure 2 In children, the center point of the femoral head is considered the ossification nucleus of the femoral head. If the ossification nucleus does not exist, the midpoint of the uppermost edge of the femur is used.

[0031] b. Graph Convolutional Network: The input to the graph convolutional network M2 consists of grouped keypoints obtained by the point detection network and image features from the first-stage coarse classification network M1. Structural keypoints (such as the highest point of the iliac crest, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity) are defined as graph nodes, and the adjacency relationships between nodes are defined as edges. A graph convolutional network (GCN) is used to learn the embedded representations of the nodes (including node feature information, adjacency relationships, multi-hop propagation, etc., see...). Figure 3 The image features of the coarse-to-medium classification network coding layer are obtained by fusing the features (using methods such as concatenation and weighted summation), and finally the image features are obtained by passing the image features through a classifier (such as a fully connected layer or Softmax).

[0032] 4) Training and inference of the DDH classification model During training, multi-stage classification models can be trained end-to-end directly, or the first stage model can be trained first, and its parameters frozen after stabilization before training the second stage classification model. During inference, given data, the corresponding classification result can be output end-to-end.

[0033] A second aspect of the present invention provides a multi-subgroup hip dysplasia classification system using the above-described multi-subgroup hip dysplasia classification method.

[0034] While various embodiments of the invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and intent of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of the invention. The appended claims are intended to define the scope of protection of the invention and therefore cover modular compositions, equivalents, or alternatives within the scope of these claims.

Claims

1. A multi-subgroup classification method for developmental dysplasia of the hip, characterized in that, include: Acquire anteroposterior X-ray images of the patient's pelvis, and preprocess the anteroposterior X-ray images to obtain preprocessing results in order to exclude the influence of non-hip dysplasia cases; A coarse classification network is constructed using a lightweight model to extract common features from images to obtain a coarse classification result that roughly identifies the category of hip dysplasia. The common features include contour and texture. A fine classification network is constructed based on graph convolutional networks. The classification results are refined by adding additional anatomical key point expression information on the coarse classification results. The anatomical key points include the highest point of the ilium, the outermost edge of the acetabulum, the center point of the femoral head, the teardrop point, and the lowest point of the ischial tuberosity. Output detailed classification network results for doctors to analyze and use; The last layer of the coarse classification network uses the Softmax function to transform the output into a probability distribution. The maximum value in the output probability distribution is the confidence score of the classification result. Data with confidence levels below a set threshold are set to enter the subsequent sub-classification stage for further refinement; data with confidence levels above a set threshold are entered into the subsequent sub-classification according to a set proportion to ensure the generalization of the sub-classification model. Based on the preprocessing results, the key anatomical points on the femoral head, acetabulum, and pelvis are obtained through a neural network. In children, the center point of the femoral head is identified as the ossification nucleus of the femoral head. If the ossification nucleus does not exist, the midpoint of the uppermost edge of the femur is used. The fine classification network defines the anatomical key points as graph nodes, and the adjacency relationship between nodes as edges; Graph convolutional networks are used to learn the embedded representations of nodes, including node feature information, adjacency relationships, and multi-hop propagation. The feature representations between key points are obtained, and then the image features of the coding layer in the coarse classification network are fused together. Finally, the image features are evaluated by a classifier for fine classification.

2. The multi-subgroup hip dysplasia classification method according to claim 1, characterized in that: The coarse classification network uses lightweight models such as MobileNet and SqueezeNet to ensure fast processing speed.

3. The multi-subgroup hip dysplasia classification method according to claim 2, characterized in that: The relevant anatomical structure points on the femoral head, acetabulum, and pelvis are obtained through neural networks, including using Faster R-CNN, UNet-like networks, or Transform-like networks.

4. The multi-subgroup hip dysplasia classification method according to claim 1, characterized in that: The fusion is performed using either concatenation or weighted summation. The classifier includes a fully connected layer or a Softmax layer.

5. The multi-subgroup hip dysplasia classification method according to claim 1, characterized in that: The image preprocessing includes excluding the following conditions: hip dysplasia that has undergone surgical treatment, combined with other hip joint diseases, pelvic internal or external rotation at the time of imaging, and pelvic, iliac, or femoral incompleteness.

6. The multi-subgroup hip dysplasia classification method according to claim 5, characterized in that: The image preprocessing also includes denoising, window width / window level adjustment, and histogram equalization of the received image to optimize image quality.

7. A multi-subgroup classification system for developmental dysplasia of the hip, characterized in that, Use the multi-subgroup hip dysplasia classification method as described in any one of claims 1-6.

Images