Ankle joint talus cartilage segmentation and thickness measurement method
Through deep learning technology, the construction of an ankle talus osteocartilage segmentation network model has solved the problems of inefficient and subjective error in the measurement of ankle talus osteocartilage thickness in the existing technology, achieving more accurate and efficient cartilage thickness measurement, and providing faster health Evaluation tool.
Patent Information
- Application Number
- CN202411095970.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-12
AI Technical Summary
In the prior art, the measurement of the thickness of the ankle talus osteocartilage relies on manual operation, is inefficient and subjective errors, and the cartilage thickness has no direct correlation with the shape of the bone, making it difficult to predict by bone CT or X-ray.
Using deep learning technology, ankle talus osteocartilage segmentation network model is constructed, through pre-processing, labeling and data enhancement of nuclear magnetic image data, multi-scale depth separable convolution blocks and KAN-enhanced multi-scale upsampling blocks for cartilage segmentation, combined with Dice Loss and binary cross entropy loss are optimized to calculate the average thickness of the cartilage.
It significantly improves the accuracy and efficiency of cartilage segmentation of ankle joints, provides doctors with faster and more accurate cartilage health assessment tools, and improves the diagnostic process and treatment effect.
Smart Images

Figure CN119006414A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a method for segmenting and measuring the thickness of talar cartilage of an ankle joint. Background Art
[0002] The health of the talar cartilage in the ankle joint is crucial in clinical diagnosis and treatment. Measuring the thickness of the talar cartilage can not only help orthopedic surgeons predict the amount of cartilage that needs to be repaired, but also avoid overtreatment and reduce harm to patients.
[0003] In the existing technology, ultrasound and magnetic resonance imaging (MRI) are commonly used methods for cartilage measurement. Both are radiation-free and safe, but studies have shown that MRI is superior to ultrasound in the accuracy of real cartilage measurement, making it the best choice for indirect observation of cartilage.
[0004] The current method of measuring cartilage thickness mainly relies on manual operation, which has problems of low efficiency and subjective errors. In addition, studies have found that cartilage thickness has no direct correlation with the length and thickness of bones, so cartilage thickness cannot be simply predicted through bone CT or X-rays.
[0005] In recent years, deep learning technology has shown great potential in the field of medical image processing, especially in image segmentation and feature extraction. The semantic segmentation method based on deep learning can automatically learn and identify cartilage tissue in images based on a large amount of labeled data, significantly improving the accuracy and efficiency of segmentation. However, there has been no systematic research report on the measurement of ankle cartilage thickness in China. Summary of the invention
[0006] In view of the deficiencies in the prior art, the object of the present invention is to provide a method for segmenting and measuring the thickness of the talar cartilage of the ankle joint.
[0007] To achieve the above object, the present invention provides the following technical solutions:
[0008] A method for segmenting and measuring the thickness of talar cartilage of an ankle joint comprises the following steps:
[0009] S1, collecting MRI images of the talus of the human ankle joint and performing data preprocessing;
[0010] First, the MRI images of the ankle joint and talus were exported from the hospital. The MRI images of the ankle joint and talus used were plain scan images of the ankle joint with an original resolution of 512′512 and the image format was jpg.
[0011] Then, the talar cartilage layer in the MRI images was screened, and the MRI images with clearly visible talar cartilage were retained; the MRI images of the talus of the ankle joint used were limited by the hospital equipment and shooting environment, and some MRI images could not clearly show the talar cartilage layer of the ankle joint, so they needed to be screened manually to retain the MRI images with clearly visible talar cartilage;
[0012] Finally, the MRI images were uniformly cropped to retain the talar cartilage and place the talar cartilage in the center of the image. The image resolution used to create the ankle talar cartilage segmentation dataset was 160′160 and the image format was png.
[0013] S2, label the segmentation labels of the talar cartilage of the ankle joint, and perform data enhancement and data set division;
[0014] First, the talar cartilage segmentation label of the ankle joint is annotated according to the contour of the talar cartilage, and a single-channel binary image containing the talar cartilage area of the ankle joint is generated. A single channel means that the image has only one grayscale channel, and the value range of each pixel is usually an integer between 0 and 255, which represents the brightness or grayscale value of the pixel. Therefore, each pixel uses a value to represent its position on the grayscale. Binarization means that each pixel has only two possible values, corresponding to the presence and absence of a certain attribute in the image. Such an image can simplify analysis and processing, and this image is used as the label of the talar cartilage segmentation dataset; the pixel value of the foreground talar cartilage area is 1, and the pixel value of the background non-talar cartilage area is 0;
[0015] Then, a variety of data enhancement methods are used to expand the number of samples in the ankle talar cartilage segmentation dataset;
[0016] The data enhancement method includes: mirror flipping along the vertical axis of the image, that is, flipping the image along the vertical axis, which helps the model learn the performance of objects in different directions; random -15° to +15° angle rotation, which enhances the model's adaptability to object rotation changes, so that it can better handle the performance of objects at different angles; random 0.8 to 1.2 times scale scaling, randomly selecting a scale ratio between 0.8 and 1.2, so that the model can handle objects at different scales; scaling the image, shearing transformation, that is, stretching a part of the image along the horizontal or vertical direction to form a parallelogram, which can simulate the deformation of the object when the viewing angle changes; Gaussian noise, adding random noise that obeys Gaussian distribution to the pixel values of the image. By introducing noise, the model is more likely to learn the characteristics of real-world data and improves generalization ability; Gaussian blur, Gaussian blur processing is performed on the image, and the blur degree is controlled by a standard deviation parameter, which can reduce details and noise in the image and help the model focus on the overall shape and structure of the object;
[0017] Finally, the training set and test set of the ankle talar cartilage segmentation dataset were randomly divided into two sets with a ratio of 7:3 in the number of images.
[0018] S3. Build a network model for segmenting the talar cartilage of the ankle joint and use the labeled data to train the model;
[0019] First, a feature extraction part in the network model is constructed; the feature extraction part is composed of a multi-scale depth-separable convolution block (MDSC block); the feature extraction part is composed of five stages, generating from lower-level abstract features to higher-level semantic features;
[0020] The core component of the feature extraction part is the multi-scale depthwise separable convolution block (MDSC block). First, the number of input channels is adjusted through the maximum pooling layer. The operation of the maximum pooling layer can basically be understood as selecting the maximum value as the output in each local area of the input (such as a 2′2 window), which is used to reduce the spatial size of the feature map while retaining important features; then, features are extracted through multi-scale depthwise separable convolution. Depthwise separable convolution is a special type of convolution operation in convolutional neural networks (CNNs), which can effectively reduce the number of model parameters and computational complexity, while improving the performance of the model to a certain extent. The convolution is composed of five depthwise convolutions of different sizes to extract features in parallel. After the extracted features are fused, feature fusion in deep learning refers to combining feature information from different levels or different sources to produce a richer and more expressive feature representation; then point-by-point convolution is used to align the number of channels; then, the batch normalization layer normalizes the output of the multi-scale depthwise separable convolution. Batch normalization is a commonly used technique in deep learning to accelerate the training process of neural networks and improve the convergence speed and stability of the model; finally, ReLU is used as the activation function;
[0021] The ReLU activation function is defined by formula (1) as:
[0022] f(x)=max(x,0) (1)
[0023] Then, constructing a talar cartilage segmentation part; the talar cartilage segmentation part is composed of a KAN enhanced multi-scale upsampling block (KEMU block); the talar cartilage segmentation part is used to generate a pixel-level talar cartilage segmentation result;
[0024] The talar cartilage segmentation part fuses the upsampled feature map with the feature map of the corresponding level through skip connection, maps the extracted multi-level features back to the original image size layer by layer, and generates pixel-level segmentation results by the last layer of the talar cartilage segmentation network. The composition of the talar cartilage segmentation part is similar to that of the feature extraction network, except that the KAN linear layer and channel spatial attention in the Kolmogorov-Arnold Network (KAN) are introduced. The KAN linear layer, as an enhanced linear layer, provides stronger nonlinear representation capabilities; the channel spatial attention enhances the feature map by gradually applying channel attention and spatial attention;
[0025] S4, construct the loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters;
[0026] First, initialize the model, load the dataset and perform data preprocessing. The dataset uses the Pascal VOC (Visual Object Classes) format; data preprocessing includes randomly adjusting the image size to the specified range (200′200-300′300), horizontally flipping, randomly cropping the image, and standardizing the pixel values of the image. The pixel values of the image are adjusted according to certain rules so that each part of the image has similar data distribution or statistical characteristics; the feature extraction part extracts the multi-level features of the original ankle talus MRI image, and the talar cartilage segmentation part segments the talar cartilage to obtain the talar cartilage segmentation prediction map;
[0027] Then, the talar cartilage segmentation loss is calculated, that is, the loss between the talar cartilage segmentation prediction map and the corresponding talar cartilage segmentation label;
[0028] The loss is the sum of Dice loss and binary cross entropy loss, which is defined by formula (2):
[0029]
[0030] In formula (2), gi is the true label of each talar cartilage segmentation, is the corresponding talar cartilage segmentation prediction image, M is the total number of images in the training set, and e is a very small normal number to ensure that the denominator is not 0;
[0031] Finally, the gradient is calculated through the back-propagation algorithm, and the optimizer is used to update the parameters of the model;
[0032] The optimizer is the AdamW optimizer, and its parameter update strategy is defined by formula (3):
[0033]
[0034] In formula (3), qt is the parameter at the tth iteration, and are the first-order and second-order momentum of the gradient, h is the learning rate, set to 0.001, e is a very small constant, set to 1e-8, and λ is the weight decay coefficient, set to 0.01;
[0035] S5, loading the trained weight parameters into the constructed segmentation model, inputting the test set data into the model, performing image post-processing on the output image, and obtaining the segmentation result of the talar cartilage;
[0036] First, load the trained model weight parameters and input the test set data into the model; then, perform image post-processing on the output image and convert the output binary image into an RGB image; finally, obtain the segmentation result image of the talar cartilage;
[0037] S6. Calculate the average thickness of talar cartilage using the segmentation results;
[0038] First, load the segmentation result image of the talar cartilage and process the image into a binary image; the segmentation result image is a grayscale image, and the grayscale image is converted into a binary image. This process will set the pixels with grayscale values greater than the threshold to white (255) and the pixels less than or equal to the threshold to black (0) according to the threshold;
[0039] Then, calculate the conversion coefficient between pixels and actual distance; use the centimeter ruler in the original image to calculate the number of pixels per centimeter, and use the conversion coefficient to convert to the actual distance. The original image size is 512′512 pixels. The measured image is 160′160 pixels, so it is necessary to calculate the proportional relationship between the actual distance and pixels according to the scaling ratio, and convert the pixel value to the actual physical distance value;
[0040] Next, contours are detected in the binary image; contours represent continuous white areas in the image, and each contour may correspond to an area of talar cartilage;
[0041] Finally, the average thickness of the two talar cartilages in the contour area is calculated and converted into actual thickness values. For each detected contour, a mask image of the same size as the image is created. First, fill the contour with white and the rest with black. Next, traverse the mask image row by row and calculate the number of white pixels in each row, which represents the thickness of the talar cartilage. By averaging the thickness of all rows, the average thickness value of this area is obtained.
[0042] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:
[0043] (1) The present invention provides a unique dataset specifically designed for ankle talar cartilage segmentation, which solves the problem of insufficient existing data and provides the necessary resources and foundation for future related research and applications.
[0044] (2) The present invention provides a network model for segmenting ankle talar cartilage, which significantly improves the accuracy and efficiency of the ankle cartilage segmentation task.
[0045] (3) The present invention provides a method for measuring the thickness of the talar cartilage of the ankle joint, providing doctors with a faster and more accurate tool for assessing cartilage health. This technological innovation helps improve the diagnostic process and enhance the treatment effect and health management level of patients. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 is a flow chart of the present invention;
[0047] Figure 2 This is a schematic diagram of the segmentation network model framework of the present invention;
[0048] Figure 3 It is a multi-scale depth-separable convolution block (MDSC block) in the segmentation network model of the present invention;
[0049] Figure 4 It is the KAN enhanced multi-scale upsampling block (KEMU block) in the segmentation network model of the present invention;
[0050] Figure 5 The test results of talar cartilage segmentation using the ankle joint talar cartilage segmentation network model in the embodiment of the present invention and the results of measuring the thickness of the ankle joint talar cartilage using the segmentation result image, wherein (a) is the input ankle joint talar cartilage segmentation dataset test image, (b) is the corresponding talar cartilage segmentation true label, (c) is the corresponding talar cartilage segmentation test mask image, (d) is the average thickness measured by the corresponding talar cartilage segmentation true label, and (e) is the average thickness measured by the corresponding talar cartilage segmentation test mask image. DETAILED DESCRIPTION
[0051] Reference Figures 1 to 5 The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.
[0052] A method for segmenting and measuring the thickness of the talar cartilage of the ankle joint, such as Figure 1 As shown, the following steps are included:
[0053] Step 1: Collect MRI images of the talus of the human ankle joint and perform data preprocessing;
[0054] Specifically, the MRI image of the ankle joint talus is exported from the hospital. The MRI image of the ankle joint talus used in this embodiment is an MR plain scan image of the ankle joint with an original resolution of 512′512, and the image format is jpg;
[0055] Screen the talar cartilage layer in the MRI image and retain a clear MRI image. The talar MRI images of the ankle joint used in this embodiment are limited by the hospital equipment and shooting environment. Some MRI images cannot clearly show the talar cartilage layer of the ankle joint, and need to be manually screened to retain the MRI images with clearly visible talar cartilage;
[0056] The present embodiment is a method for segmenting and measuring the thickness of the talar cartilage of the ankle joint. The original talar MRI image of the ankle joint has other interferences besides the talar cartilage, so it is necessary to uniformly crop and process the image, retain the talar cartilage and make the talar cartilage in the center of the image. The image resolution used to prepare the ankle joint talar cartilage segmentation data set is 160′160, and the image format is png;
[0057] Step 2: Label the ankle talar cartilage segmentation labels, and use data enhancement to expand the number of ankle MRI images to construct an ankle talar cartilage segmentation dataset; the ankle talar cartilage segmentation dataset consists of ankle MRI images and corresponding talar cartilage segmentation labels;
[0058] The details are as follows:
[0059] The ankle joint talar cartilage segmentation label is annotated according to the contour of the talar cartilage, and a single-channel binary image containing the ankle joint talar cartilage area is generated. This image is used as the label of the talar cartilage segmentation dataset; the pixel value of the foreground talar cartilage area is 1, and the pixel value of the background non-talar cartilage area is 0;
[0060] A variety of data enhancement methods are used to expand the number of samples in the ankle talar cartilage segmentation dataset, including mirror flipping along the vertical axis of the image, random -15° to +15° angle rotation, random 0.8 to 1.2 times scale scaling, shear transformation, Gaussian noise, and Gaussian blur.
[0061] The training set and test set of the ankle talar cartilage segmentation dataset were randomly divided into two sets in a ratio of 7:3.
[0062] Step 3: Build a network model for segmenting the ankle talar cartilage and use the labeled data to train the model. The network model framework is as follows: Figure 2 As shown, specifically:
[0063] Constructing a feature extraction part in a network model; the feature extraction part is composed of a multi-scale depth-separable convolution block (MDSC block); the feature extraction part is composed of five stages, generating from lower-level abstract features to higher-level semantic features;
[0064] The core component of the feature extraction part in this embodiment is a multi-scale depth-separable convolution block (MDSC block), which has a structure as follows: Figure 3 As shown in the figure. First, the number of input channels is adjusted through the maximum pooling layer; then, features are extracted through multi-scale depthwise separable convolution, which extracts features in parallel by five depthwise convolutions of different sizes. After the extracted features are fused, the number of channels is aligned using point-by-point convolution; then, the batch normalization layer normalizes the output of the multi-scale depthwise separable convolution; finally, ReLU is used as the activation function;
[0065] The ReLU activation function is defined by formula (1) as:
[0066] f(x)=max(x,0) (1)
[0067] Construct a talar cartilage segmentation part; the talar cartilage segmentation part is composed of a KAN enhanced multi-scale upsampling block (KEMU block), and the structure is as follows Figure 4 As shown. The talar cartilage segmentation part is used to generate pixel-level talar cartilage segmentation results;
[0068] In this embodiment, the talar cartilage segmentation part fuses the upsampled feature map with the feature map of the corresponding level through jump connections, maps the extracted multi-level features back to the original image size layer by layer, and generates pixel-level segmentation results by the last layer of the talar cartilage segmentation network. Part of the composition of the talar cartilage segmentation part is similar to that in the feature extraction network, except that the KAN linear layer and channel spatial attention in the Kolmogorov-Arnold Networks (KAN) are introduced. As an enhanced linear layer, the KAN linear layer provides stronger nonlinear representation capabilities; the channel spatial attention enhances the feature map by gradually applying channel attention and spatial attention, so that the network can pay more attention to important features;
[0069] Step 4: Construct the loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters;
[0070] The details are as follows:
[0071] Initialize the model, load the dataset and perform data preprocessing;
[0072] The dataset uses the Pascal VOC (Visual Object Classes) format, which is highly adaptable to various network models. Data preprocessing includes randomly adjusting the image size to a specified range (200′200-300′300), horizontal flipping, randomly cropping the image, and standardizing the pixel value of the image.
[0073] The feature extraction network is used to extract the multi-level features of the original ankle joint talus MRI image, and the talus cartilage segmentation network is used to segment the talus cartilage to obtain the talus cartilage segmentation prediction map;
[0074] Calculate the talar cartilage segmentation loss, that is, the loss between the talar cartilage segmentation prediction map and the corresponding talar cartilage segmentation label;
[0075] The loss is the sum of Dice loss and binary cross entropy loss, which is defined by formula (2):
[0076]
[0077] In formula (2), gi is the true label of each talar cartilage segmentation, is the corresponding talar cartilage segmentation prediction image, M is the total number of images in the training set, and e is a very small normal number to ensure that the denominator is not 0;
[0078] The gradient is calculated through the back-propagation algorithm and the parameters of the model are updated using the optimizer;
[0079] The optimizer is the AdamW optimizer, and its parameter update strategy is defined by formula (3):
[0080]
[0081] In formula (3), qt is the parameter at the tth iteration, and are the first-order and second-order momentum of the gradient, h is the learning rate, set to 0.001, e is a very small constant, set to 1e-8, and λ is the weight decay coefficient, set to 0.01;
[0082] Step 5: The constructed segmentation model is loaded with the weight parameters trained above, the test set data is input into the model, and the output image is post-processed to obtain the segmentation result of the talar cartilage;
[0083] Specifically, the trained model weight parameters are loaded, and the test set data is input into the model; the output image is post-processed, and the output binary image is converted into an RGB image; the segmentation result image of the talar cartilage is obtained, as shown in Figure 5As shown, (a) is the input test image of the ankle joint talar cartilage segmentation dataset, (b) is the corresponding talar cartilage segmentation true label, and (c) is the corresponding talar cartilage segmentation test mask image;
[0084] Step 6: Use the segmentation results to calculate the average thickness of the talar cartilage;
[0085] The details are as follows:
[0086] Load the segmentation result image of the talar cartilage and convert it into a binary image. This process will set the pixels with grayscale values greater than the threshold to white (255) and the pixels less than or equal to the threshold to black (0).
[0087] Calculate the conversion coefficient between pixels and actual distance, use the centimeter ruler in the original image to calculate the number of pixels per centimeter, and use the conversion coefficient to convert to the actual distance. The original image size is 512′512 pixels. The measured image is 160′160 pixels, so it is necessary to calculate the proportional relationship between the actual distance and pixels according to the scaling ratio, and convert the pixel value to the actual physical distance value;
[0088] Detect contours in the binary image. Contours represent continuous white areas in the image. Each contour may correspond to an area of talar cartilage.
[0089] Calculate the average thickness of the two pieces of talar cartilage in the contour area and convert it into the actual thickness value. For each detected contour, create a mask image of the same size as the image. First fill the contour with white and the rest with black. Next, traverse the mask image row by row and calculate the number of white pixels in each row, which represents the thickness of the talar cartilage. By taking the average of the thickness of all rows, the average thickness value of this area is obtained;
[0090] like Figure 5 As shown, (d) is the average thickness measured by the corresponding talar cartilage segmentation true label, and (e) is the average thickness measured by the corresponding talar cartilage segmentation test mask.
[0091] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.
Claims
1. A method for segmenting and measuring the thickness of talar cartilage of an ankle joint, characterized in that: The following steps are involved: S1, collecting MRI images of the talus of the human ankle joint and performing data preprocessing; S2, label the segmentation labels of the talar cartilage of the ankle joint, and perform data enhancement and data set division; S3. Build a network model for segmenting the talar cartilage of the ankle joint and use the labeled data to train the model; S4, construct the loss function, determine the optimizer and hyperparameters, input the training set, and train the network parameters; S5, loading the trained weight parameters into the constructed segmentation model, inputting the test set data into the model, performing image post-processing on the output image, and obtaining the segmentation result of the talar cartilage; S6. Calculate the average thickness of the talar cartilage using the segmentation results.
2. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 1, characterized in that: Wherein step S1 comprises the following: First, the MRI images of the ankle joint and talus were exported from the hospital. The MRI images of the ankle joint and talus used were plain scan images of the ankle joint with an original resolution of 512′512 and the image format was jpg. Then, screen the talar cartilage layer in the MRI image and retain the clear MRI image. The MRI images of the ankle talus used are limited by the hospital equipment and shooting environment. Some MRI images cannot clearly show the talar cartilage layer of the ankle joint. Manual screening is required to retain the MRI images with clear visible talar cartilage. Finally, the MRI images were uniformly cropped to retain the talar cartilage and place it in the center of the image. The image resolution used to prepare the ankle talar cartilage segmentation dataset was 160′160 and the image format was png.
3. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 2, characterized in that: Wherein step S2 comprises the following: Firstly, the segmentation label of the talar cartilage of the ankle joint is annotated according to the contour of the talar cartilage, and a single-channel binary image containing the talar cartilage area of the ankle joint is generated, and this image is used as the label of the talar cartilage segmentation dataset; The pixel value of the foreground talar cartilage area is 1, and the pixel value of the background non-talar cartilage area is 0; Then, a variety of data enhancement methods are used to expand the number of samples in the ankle talar cartilage segmentation dataset; the data enhancement methods include: mirror flipping along the vertical axis of the image, random -15° to +15° angle rotation, random 0.8 to 1.2 times scale scaling, shear transformation, Gaussian noise and Gaussian blur; Finally, the training set and test set of the ankle talar cartilage segmentation dataset were randomly divided into two sets in a ratio of 7:
3.
4. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 3, characterized in that: Wherein step S3 comprises the following: First, a feature extraction part in the network model is constructed; the feature extraction part is composed of a multi-scale depth-separable convolution block (MDSC block); the feature extraction part is composed of five stages, generating from lower-level abstract features to higher-level semantic features; The core component of the feature extraction part is the multi-scale depthwise separable convolution block (MDSC block). First, the number of input channels is adjusted through the maximum pooling layer; then, features are extracted through multi-scale depthwise separable convolution, which extracts features in parallel by five depthwise convolutions of different sizes, fuses the extracted features, and then uses point-by-point convolution to align the number of channels; then, the batch normalization layer normalizes the output of the multi-scale depthwise separable convolution; finally, ReLU is used as the activation function; The ReLU activation function is defined by formula (1) as: f(x)=max(x,0) (1) Then, constructing a talar cartilage segmentation part; the talar cartilage segmentation part is composed of a KAN enhanced multi-scale upsampling block (KEMU block); the talar cartilage segmentation part is used to generate a pixel-level talar cartilage segmentation result; The talar cartilage segmentation part uses skip connections to fuse the upsampled feature map with the feature map of the corresponding level, maps the extracted multi-level features back to the original image size layer by layer, and generates pixel-level segmentation results by the last layer of the talar cartilage segmentation network; The composition of the talar cartilage segmentation part is similar to that of the feature extraction network, except that the KAN linear layer and channel spatial attention in the Kolmogorov-Arnold Network (KAN) are introduced. The KAN linear layer, as an enhanced linear layer, provides stronger nonlinear representation capabilities; Channel-Spatial Attention enhances feature maps by progressively applying channel-wise attention and spatial attention.
5. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 4, characterized in that: Wherein step S4 comprises the following: First, initialize the model, load the dataset and perform data preprocessing. The dataset uses the Pascal VOC (Visual Object Classes) format; data preprocessing includes randomly adjusting the image size to the specified range (200′200-300′300), horizontally flipping, randomly cropping the image, and standardizing the pixel value of the image; the feature extraction part extracts the multi-level features of the original ankle joint talus MRI image, and the talar cartilage segmentation part segments the talar cartilage to obtain the talar cartilage segmentation prediction map; Then, the talar cartilage segmentation loss is calculated, that is, the loss between the talar cartilage segmentation prediction map and the corresponding talar cartilage segmentation label; The loss is the sum of Dice loss and binary cross entropy loss, which is defined by formula (2): In formula (2), gi is the true label of each talar cartilage segmentation, is the corresponding talar cartilage segmentation prediction image, M is the total number of images in the training set, and e is a very small normal number to ensure that the denominator is not 0; Finally, the gradient is calculated through the back-propagation algorithm, and the optimizer is used to update the parameters of the model; The optimizer is the AdamW optimizer, and its parameter update strategy is defined by formula (3): In formula (3), qt is the parameter at the tth iteration, and are the first-order and second-order momentum of the gradient, h is the learning rate, e is a very small constant, and λ is the weight decay coefficient.
6. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 5, characterized in that: Wherein step S5 comprises the following: First, the trained model weight parameters are loaded and the test set data is input into the model; then, the output image is post-processed and the output binary image is converted into an RGB image; finally, the segmentation result image of the talar cartilage is obtained.
7. The method for segmenting and measuring the thickness of the talar cartilage of the ankle joint according to claim 6, characterized in that: Wherein step S6 comprises the following: First, load the segmentation result image of the talar cartilage and process the image into a binary image; the segmentation result image is a grayscale image, and the grayscale image is converted into a binary image. This process will set the pixels with grayscale values greater than the threshold to white (255) and the pixels less than or equal to the threshold to black (0) according to the threshold; Then, calculate the conversion coefficient between pixels and actual distance; use the centimeter ruler in the original image to calculate the number of pixels per centimeter, and use the conversion coefficient to convert to the actual distance. The original image size is 512′512 pixels. The measured image is 160′160 pixels, so it is necessary to calculate the proportional relationship between the actual distance and pixels according to the scaling ratio, and convert the pixel value to the actual physical distance value; Next, contours are detected in the binary image; contours represent continuous white areas in the image, and each contour may correspond to an area of talar cartilage; Finally, the average thickness of the two talar cartilages in the contour area is calculated and converted into actual thickness values. For each detected contour, a mask image of the same size as the image is created. First, fill the contour with white and the rest with black. Next, traverse the mask image row by row and calculate the number of white pixels in each row, which represents the thickness of the talar cartilage. By averaging the thickness of all rows, the average thickness value of this area is obtained.
Citation Information
Patent Citations
Cartilage thickness detection method and device, computer device and readable storage medium
CN110827285A
Industrial CT image segmentation method based on deep convolutional neural network
CN112330682A
Knee joint nuclear magnetic resonance image automatic segmentation method based on deep learning
CN115953416A
Concrete chloride ion diffusion coefficient prediction method based on aggregate grading
CN116559045A
Oracle bone font classification method based on deep learning
CN117333887A
Cited By
Bone injury degree evaluation system based on artificial intelligence
CN120183679A
Enhanced CT (Computed Tomography) image generation method and equipment based on double-space constraint
CN120876659A
