MRI image segmentation method based on WKA-Unet

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging) image segmentation method based on WKA-Unet. The MRI image segmentation method comprises the following steps: 1) data preprocessing; 2) constructing a WKA-Unet network; and 3) testing. According to the technical scheme, on the basis of a U-Net network structure, an MRI image segmentation model is improved; a traditional Unet network is combined with a self-designed WKA module, multi-scale features are fully captured by using wavelet transform convolution and KAN convolution, and the loss of traditional Unet global information is relieved; an attention mechanism is introduced to pay attention to key parts, and redundant information is removed; according to the technical scheme, the network is trained more flexibly and accurately, a more accurate segmentation result is obtained, the segmentation precision is high, and the model structure is efficient and stable.

Description

Technical Field

[0001] The invention relates to image processing technology, in particular to an MRI image segmentation method based on WKA-Unet. Background Art

[0002] Image segmentation is an important task in the field of computer vision, which aims to divide an image into different regions with similar features. Early image segmentation algorithms were mainly based on traditional computer vision techniques, including threshold processing, region growing, edge detection, etc. Michael Kass et al. proposed the active contour models image segmentation method in 1988, using a curve that can automatically shrink and expand. The curve can move around the features perceived in the image. The shape of this curve is determined by minimizing an energy function. By optimizing this energy function, a curve that adapts to the edge of the image can be obtained. R.Adams et al. proposed a region growing method for image segmentation. Through the merging and growing process, the algorithm can adaptively generate image regions with similar characteristics. It is often used in image segmentation, especially for locating and segmenting organs and lesion structures. John Canny et al. proposed an edge detection algorithm to solve the problem of accurately detecting edges in images. The Canny edge detection algorithm not only performs well in practice, but also has been fully explained in theory. It is an efficient and robust edge detection method in the field of image processing. These methods perform well in some scenarios, but are usually limited in complex backgrounds and large changes. With the rise of deep learning, convolutional neural networks (CNNs) and other deep learning architectures have achieved remarkable results in image segmentation: Imran et al. proposed a new progressive adversarial semantic segmentation model that can make improved segmentation predictions during training without the need for data in a specific field; Ronneberger et al. proposed the U-Net network, which consists of an encoder, a decoder, and a jump connection. The encoder performs multiple downsampling and convolution operations to reduce the image resolution while obtaining image context information. The decoder first restores the image resolution through multiple upsampling operations, then fuses the underlying information obtained by the encoder with the restored high-level information through jump connections, and finally extracts the fused features through multiple convolution operations. U-Net has achieved good results in image segmentation tasks and has been widely used. In order to meet the actual needs of different tasks, different scholars have improved U-Net and proposed a series of improved networks. Ozan Oktay et al. introduced the attention mechanism based on U-Net. The attention mechanism allows the network to focus more on important areas in the image, thereby improving the segmentation performance of key structures.The overall structure of Attention U-Net is still based on the encoder-decoder structure of U-Net. On the basis of U-Net, Attention U-Net introduces the attention gate to adjust the weight of the features so that the network can pay more attention to the area of ​​interest; Zongwei Zhou et al. pointed out that although U-Net performs well in segmentation tasks, its modeling ability of multi-scale information is relatively limited. For this reason, they proposed an improved model UNet++, which is improved through a nested and dense U-Net structure. In UNet++, each U-Net module is designed as a "sub-network" and nested in the main U-Net network to form a multi-level structure.

[0003] The application of deep learning in image segmentation has become a general trend. With great breakthroughs in computer performance, deep learning has developed rapidly and achieved great success in the field of image processing. The continuous pooling and convolution of the U-Net network will reduce the resolution of the image, resulting in the loss of some spatial information, making it difficult to extract more features and resulting in low segmentation accuracy. There are more and more improvement schemes. Summary of the invention

[0004] The purpose of the present invention is to provide an MRI image segmentation method based on WKA (Wave-KAN-Attention, abbreviated as WKA)-Unet in view of the shortcomings of the prior art. This method can improve the segmentation performance of the model for small targets, and can segment the edge details more accurately for MRI images with multiple tissue overlaps and interferences.

[0005] The technical solution for achieving the purpose of the present invention is:

[0006] A WKA-Unet-based MRI image segmentation method comprises the following steps:

[0007] 1) Data preprocessing: Collect image data sets. Since the data sets come from different locations, time points, and imaging devices, individual differences in data and data heterogeneity are inevitable. At the same time, image quality also faces challenges, such as uneven data distribution, significant contrast fluctuations, and noise interference. Therefore, before image segmentation, a series of systematic preprocessing steps must be performed. Data preprocessing includes abnormal image processing, Gaussian noise, size scaling, contrast adjustment, and data enhancement. The purpose of data preprocessing is to improve image quality and quantity, ensure data standardization, and thus improve the accuracy and credibility of subsequent segmentation algorithms. Specifically:

[0008] 1-1) Abnormal image processing: In order to ensure the quality and consistency of the data, it is necessary to remove abnormal images that are not suitable for training to improve the segmentation accuracy of the model. Abnormal images include: images with severe imaging artifacts, low-resolution and low-quality images, images that do not contain the complete target area, and images with severe lesions. Abnormal images are not suitable for model training. Images with too low resolution, too low quality, and too high noise will cause blurred image edges, making it impossible for the network to accurately learn the boundary features of the image. Ensuring that the image resolution can be clearly distinguished is a key step in image preprocessing. Through abnormal image processing, image samples with poor quality or blur are removed to ensure the overall reliability of the experimental data.

[0009] 1-2) Gaussian noise: Add Gaussian noise to the image to simulate MRI images in a noisy environment. Generate random numbers that conform to the Gaussian distribution through a random number generator, add the generated random values ​​to each pixel of the image, and use the numpy.random.normal() function in Numpy to generate noise that satisfies the Gaussian distribution and superimpose it on the image pixel value;

[0010] 1-3) Scaling: Use the resize() function of the image processing library or programming language to adjust the image size. When the nearest neighbor interpolation method is used, for each target pixel, its value will be set to the value of the nearest input pixel. Specifically: adjust the image size and scale the input image to the specified target size; assign pixel values. For each target pixel, determine the input pixel closest to it and assign the value of the input pixel to the target pixel. The image is finally uniformly cropped or scaled to 256×256 pixels to eliminate the impact caused by size differences.

[0011] 1-4) Contrast adjustment: Use image contrast adjustment, i.e. histogram equalization and contrast stretching, to enhance the contrast of the image and highlight the area of ​​interest. Specifically, perform histogram equalization on the image to enhance the contrast of the image; use OpenCV's cv2.equalizeHist() function to equalize the grayscale image;

[0012] 1-5) Data enhancement: Geometric transformation and flipping are used to expand the data set to overcome its size limitation. Geometric transformation rotates the MRI image 30 degrees clockwise through rotation operation. This rotation can change the angular position of the edge of the image in the image, allowing the model to learn the features of the object at different angles. Flipping is to generate a mirror image by horizontal mirroring operation, which can effectively increase the number of training samples, thereby achieving the purpose of data enhancement. At the same time, horizontal mirroring can increase the diversity of the data set, so that the model can better learn features and reduce the risk of overfitting on the training set. In MRI images, there may be multiple transformation forms, and the target may change in different directions. Mirroring can help the model better identify these changes;

[0013] 1-6) Classification: Expand the image dataset after data preprocessing in steps 1-1) to 1-5) to 1,000 images, and divide them into a training set and a test set in a ratio of 7:3;

[0014] 2) Constructing WKA-Unet network: WKA-Unet is improved based on Unet network, and the WKA module is introduced. The WKA module is embedded in each layer of the Unet network, which can strengthen the feature representation ability layer by layer, thereby improving the robustness and accuracy of the entire network when processing complex features. Among them, the Unet network is a convolutional neural network architecture for biomedical image segmentation. It introduces a novel network structure and training strategy to solve the challenges faced by traditional methods when the amount of data is insufficient. The main idea of ​​the Unet network is to use data enhancement technology to efficiently utilize limited labeled samples and improve segmentation accuracy through a unique network design. The basic structure includes:

[0015] Encoder: A series of convolutional pooling layers are used to capture the contextual information of the image. There are convolutional pooling layers for downsampling operations. Although this helps capture the contextual information of the image, pooling operations (such as maximum pooling or average pooling) will inevitably lose some spatial information. Although the upsampling layer of the encoder can try to restore the fine details of the segmentation, due to the information loss in the previous downsampling process, the recovery effect after upsampling may be limited. When restoring details, upsampling methods (such as bilinear interpolation and transposed convolution) are more based on inference based on the existing downsampled features, and it is difficult to completely restore all the detail information in the original image. Considering that traditional convolution operations mainly rely on local receptive fields, although the Unet network expands the receptive field through downsampling and upsampling, the capture of global contextual information is still limited, which may affect the understanding of the overall structure of the image. In view of the shortcomings of the Unet process due to downsampling and upsampling, the WKA wavelet feature enhancement convolution module is used to improve the Unet network;

[0016] Decoder: A series of upsampling and convolutional layers are used to recover the fine details of the segmentation;

[0017] Skip connection: copy the feature map from the contraction path to the corresponding expansion path to retain more detail information;

[0018] The WKA module is the core improvement of WKA-Unet. It combines wavelet transform convolution, KAN convolution and CBAM attention mechanism to improve the feature extraction and representation capabilities of the model. In the design of the WKA module, different convolution operations and attention mechanisms are integrated to help capture the multi-scale information of features, spatial contextual relationships and correlations between channels, including:

[0019] 2-1) Wavelet transform convolution: The input feature map X is first processed by wavelet transform convolution to obtain feature map X1. Wavelet transform convolution is the first step in the WKA module. The input feature map X is first processed by wavelet transform convolution to obtain feature map X1. Wavelet transform provides a method to decompose a signal into components that can provide time and frequency information. Therefore, wavelet transform convolution can capture information at different frequency and spatial scales by decomposing the input signal using wavelet transform. The multi-level decomposition of wavelet transform enables wavelet transform convolution to cover a wider range of input data while keeping the number of parameters relatively low, which means that even a small convolution kernel can capture a wider range of contextual information through wavelet transform. In addition, wavelet transform convolution allows the network to perform independent convolution processing on the decomposed different frequency features, which enhances the model's ability to respond to different features in the signal, especially the capture of low-frequency features, which is very important for understanding the shape and structure in MRI images. After processing the signal in the wavelet domain, the wavelet transform convolution uses the inverse wavelet transform to recombine the processed signal to generate the final output. The input feature map is decomposed into low-frequency and high-frequency parts of different scales through multiple wavelet transforms, and the shape at each scale is recorded. For the high-frequency part of each scale, after being processed by the wavelet transform convolution and scaling module, it is recombined with the low-frequency part for inverse wavelet transform and gradually restored to the original scale, thereby extracting features at different scales, especially low-frequency information;

[0020] 2-2) KAN convolution: The input feature map X is also processed by KAN convolution to obtain feature map X2. KAN convolution is based on the Kolmogorov-Arnold theorem, which provides a method to represent a multivariate continuous function as a combination of a univariate function and addition. This idea has been effectively applied in the design of Kolmogorov-Arnold networks (KANs). The introduction of KANs into the convolution layer is an innovation in this field. MRI images usually have the characteristic of high noise, and the traditional convolution used by the original Unet may cause the loss of contextual information during feature extraction. This is because the convolution operation usually only focuses on local features, and when downsampling, the larger receptive field may fail to fully capture the global contextual information, resulting in the loss of contextual information. KAN convolution is based on spline The convolution operation enhances the model's ability to capture nonlinear relationships, allowing the network to extract more complex and expressive features, especially when processing highly complex images. Secondly, its multi-scale feature fusion capability further enhances the combination of local and global information, which helps to improve the transmission of semantic information. In addition, traditional convolution applies dot products between the kernel and the corresponding pixels in the image, while KAN convolution can apply a learnable nonlinear activation function to each element and then sum it. Learnability means that it provides more expressiveness for the activation function. Compared with traditional convolution, KAN convolution can effectively reduce the number of model parameters while maintaining or improving feature extraction capabilities, thereby improving training efficiency and inference speed. These advantages make the application of KAN convolution in U-Net particularly suitable for tasks that require high-precision feature extraction, such as image segmentation.

[0021] 2-3) Fusion: The feature map X1 obtained by wavelet convolution and the feature map X2 obtained by KAN convolution are fused to generate a new feature Y. The attention mechanism is further introduced. The attention mechanism emphasizes the key feature area and suppresses redundant information by weighting. It can help the model more effectively integrate features from different convolutional layers, ensure that key information is retained and strengthened, and promote more accurate segmentation results. CBAM (Convolutional Block Attention Module, referred to as CBAM) is an attention mechanism used to improve the performance of convolutional neural networks. It enhances the expressiveness of feature maps through two key steps: channel attention and spatial attention. In the channel attention module, CBAM uses global average pooling and global maximum pooling to extract global features of each channel, and generates channel weights through a shared multi-layer perceptron (MLP) to emphasize feature channels that are important to the task. Next, the spatial attention module processes the feature map after channel weighting, generates a spatial attention map through convolution operation, and further highlights the key spatial area. Through this mechanism, CBAM effectively focuses on important features and reduces the impact of redundant information;

[0022] 2-4) Generate final features: The fused feature Y is processed by the attention mechanism CBAM to generate the final feature map Y1;

[0023] The features of wavelet transform convolution are fused with the features of KAN convolution, and finally the feature map is weighted through the attention mechanism to highlight important features and suppress irrelevant features, thereby enhancing the model's sensitivity to important information and further improving the feature expression ability. The WKA module can effectively extract global information while focusing on important features, reducing redundant information images, and fully capturing the edge features of multifidus muscle MRI images. This WKA-Unet combination strategy enables the model to have stronger capabilities when facing complex image segmentation tasks;

[0024] 3) Testing: In order to evaluate the segmentation performance of different models, a variety of evaluation indicators are used to evaluate the models, including Dice Similarity Coefficient (DSC), Jaccard Index (JAC) and Hausdorff Distance (HD). The Dice coefficient is used to quantify the similarity between the predicted segmentation area and the actual label area. The Dice coefficient value ranges from 0 to 1. The closer the value is to 1, the more consistent the segmentation result is with the actual situation. The Jaccard index is the intersection over union (IoU), which is used to measure the degree of overlap between samples. The Jaccard index value ranges from 0 to 1. The larger the value, the more obvious the overlap. The Hausdorff distance (HD) reflects the similarity between point sets. The maximum distance between the actual annotation and the predicted segmentation boundary is calculated to evaluate the segmentation accuracy. It is particularly sensitive to the accuracy of the boundary. The smaller the value, the higher the segmentation accuracy. The calculation method of the evaluation indicators is as follows: The calculation formula of DSC is:

[0025]

[0026] The calculation formula of JAC is:

[0027]

[0028] HD calculation formula:

[0029] HD(p,g)=max(h(p,g),h(g,p))

[0030]

[0031] Where p is the pixel set of the predicted segmentation, and g is the pixel set of the Ground Truth.

[0032] This method can improve the model's segmentation performance for small targets and can segment edge details more accurately for MRI images with multiple tissues overlapping and interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of a method flow chart of an embodiment;

[0034] Figure 2 It is a schematic diagram of the structure of the Unet network in the embodiment;

[0035] Figure 3 This is a schematic diagram of the WKA module structure in the embodiment;

[0036] Figure 4 A schematic diagram of characteristic changes of the WKA module of the embodiment;

[0037] Figure 5 Schematic diagram of the WKA-Unet network structure in the embodiment. DETAILED DESCRIPTION

[0038] The content of the present invention is further described below in conjunction with the drawings and embodiments, but the present invention is not limited thereto.

[0039] Example:

[0040] Refer to Figure 1 , Figure 5 , a WKA-Unet-based MRI image segmentation method comprises the following steps:

[0041] 1) Data preprocessing: Collect image data sets. Since the data sets come from different locations, time points, and imaging devices, individual differences in data and data heterogeneity are inevitable. At the same time, image quality also faces challenges, such as uneven data distribution, significant contrast fluctuations, and noise interference. Therefore, before image segmentation, a series of systematic preprocessing steps must be performed. Data preprocessing includes abnormal image processing, Gaussian noise, size scaling, contrast adjustment, and data enhancement. The purpose of data preprocessing is to improve image quality and quantity, ensure data standardization, and thus improve the accuracy and credibility of subsequent segmentation algorithms. Specifically:

[0042] 1-1) Abnormal image processing: In order to ensure the quality and consistency of the data, it is necessary to remove abnormal images that are not suitable for training to improve the segmentation accuracy of the model. Abnormal images include: images with severe imaging artifacts, low-resolution and low-quality images, images that do not contain the complete target area, and images with severe lesions. Abnormal images are not suitable for model training. Images with too low resolution, too low quality, and too high noise will cause blurred image edges, making it impossible for the network to accurately learn the boundary features of the image. Ensuring that the image resolution can be clearly distinguished is a key step in image preprocessing. Through abnormal image processing, image samples with poor quality or blur are removed to ensure the overall reliability of the experimental data.

[0043] 1-2) Gaussian noise: Add Gaussian noise to the image to simulate MRI images in a noisy environment. Generate random numbers that conform to the Gaussian distribution through a random number generator, add the generated random values ​​to each pixel of the image, and use the numpy.random.normal() function in Numpy to generate noise that satisfies the Gaussian distribution and superimpose it on the image pixel value;

[0044] 1-3) Scaling: Use the resize() function of the image processing library or programming language to adjust the image size. When the nearest neighbor interpolation method is used, for each target pixel, its value will be set to the value of the nearest input pixel. Specifically: adjust the image size and scale the input image to the specified target size; assign pixel values. For each target pixel, determine the input pixel closest to it and assign the value of the input pixel to the target pixel. The image is finally uniformly cropped or scaled to 256×256 pixels to eliminate the impact caused by size differences.

[0045] 1-4) Contrast adjustment: Use image contrast adjustment, i.e. histogram equalization and contrast stretching, to enhance the contrast of the image and highlight the area of ​​interest. Specifically, perform histogram equalization on the image to enhance the contrast of the image; use OpenCV's cv2.equalizeHist() function to equalize the grayscale image;

[0046] 1-5) Data enhancement: Geometric transformation and flipping are used to expand the data set to overcome its size limitation. Geometric transformation rotates the MRI image 30 degrees clockwise through rotation operation. This rotation can change the angular position of the edge of the image in the image, allowing the model to learn the features of the object at different angles. Flipping is to generate a mirror image by horizontal mirroring operation, which can effectively increase the number of training samples, thereby achieving the purpose of data enhancement. At the same time, horizontal mirroring increases the diversity of the data set, so that the model can better learn features and reduce the risk of overfitting on the training set. In MRI images, there may be multiple transformation forms, and the target may change in different directions. Mirroring can help the model better identify these changes;

[0047] 1-6) Classification: Expand the image dataset after data preprocessing in steps 1-1) to 1-5) to 1,000 images, and divide them into a training set and a test set in a ratio of 7:3;

[0048] 2) Construct WKA-Unet network: WKA-Unet is improved based on Unet network, and WKA module is embedded in each layer of Unet network. The improved network structure is as follows: Figure 5 As shown in Figure 1, this design can strengthen the feature representation capability layer by layer, thereby improving the robustness and accuracy of the entire network when processing complex features. Among them, the Unet network is a convolutional neural network architecture for biomedical image segmentation. It introduces a novel network structure and training strategy to solve the challenges faced by traditional methods when the amount of data is insufficient. The main idea of ​​the Unet network is to use data enhancement technology to efficiently utilize limited labeled samples and improve segmentation accuracy through a unique network design. The basic structure of Unet is as follows: Figure 2 As shown, including:

[0049] Encoder: A series of convolutional pooling layers are used to capture the contextual information of the image. There are convolutional pooling layers for downsampling operations. Although this helps capture the contextual information of the image, pooling operations (such as maximum pooling or average pooling) will inevitably lose some spatial information. Although the upsampling layer of the encoder can try to restore the fine details of the segmentation, due to the information loss in the previous downsampling process, the recovery effect after upsampling may be limited. When restoring details, upsampling methods (such as bilinear interpolation and transposed convolution) are more based on inference based on the existing downsampled features, and it is difficult to completely restore all the detail information in the original image. Considering that traditional convolution operations mainly rely on local receptive fields, although the Unet network expands the receptive field through downsampling and upsampling, the capture of global contextual information is still limited, which may affect the understanding of the overall structure of the image. In view of the shortcomings of the Unet process due to downsampling and upsampling, the WKA wavelet feature enhancement convolution module is used to improve the Unet network;

[0050] Decoder: A series of upsampling and convolutional layers are used to recover the fine details of the segmentation;

[0051] Skip connection: copy the feature map from the contraction path to the corresponding expansion path to retain more detail information;

[0052] The WKA module is the core improvement of WKA-Unet. It combines wavelet transform convolution, KAN convolution, and CBAM attention mechanism to improve the model's feature extraction and representation capabilities. In the design of the WKA module, the integration of different convolution operations and attention mechanisms helps capture the multi-scale information of features, spatial contextual relationships, and correlations between channels. The overall structure of the WKA module is as follows: Figure 3 As shown, including:

[0053] 2-1) Wavelet transform convolution: The input feature map X is first processed by wavelet transform convolution to obtain feature map X1. Wavelet transform convolution is the first step in the WKA module. The input feature map X is first processed by wavelet transform convolution to obtain feature map X1. Wavelet transform provides a method to decompose a signal into components that can provide time and frequency information. Therefore, wavelet transform convolution can capture information at different frequency and spatial scales by decomposing the input signal using wavelet transform. The multi-level decomposition of wavelet transform enables wavelet transform convolution to cover a wider range of input data while keeping the number of parameters relatively low, which means that even a small convolution kernel can capture a wider range of contextual information through wavelet transform. In addition, wavelet transform convolution allows the network to perform independent convolution processing on the decomposed different frequency features, which enhances the model's ability to respond to different features in the signal, especially the capture of low-frequency features, which is very important for understanding the shape and structure in MRI images. After processing the signal in the wavelet domain, the wavelet transform convolution uses the inverse wavelet transform to recombine the processed signal to generate the final output. The input feature map is decomposed into low-frequency and high-frequency parts of different scales through multiple wavelet transforms, and the shape at each scale is recorded. For the high-frequency part of each scale, after being processed by the wavelet transform convolution and scaling module, it is recombined with the low-frequency part for inverse wavelet transform and gradually restored to the original scale, thereby extracting features at different scales, especially low-frequency information;

[0054] 2-2) KAN convolution: The input feature map X is also processed by KAN convolution to obtain feature map X2. KAN convolution is based on the Kolmogorov-Arnold theorem, which provides a method to represent a multivariate continuous function as a combination of a univariate function and addition. This idea has been effectively applied in the design of Kolmogorov-Arnold networks (KANs). The introduction of KANs into the convolution layer is an innovation in this field. MRI images usually have the characteristic of high noise, and the traditional convolution used by the original Unet may cause the loss of contextual information during feature extraction. This is because the convolution operation usually only focuses on local features, and when downsampling, the larger receptive field may fail to fully capture the global contextual information, resulting in the loss of contextual information. KAN convolution is based on spline The convolution operation enhances the model's ability to capture nonlinear relationships, allowing the network to extract more complex and expressive features, especially when processing highly complex images. Secondly, its multi-scale feature fusion capability further enhances the combination of local and global information, which helps to improve the transmission of semantic information. In addition, traditional convolution applies dot products between the kernel and the corresponding pixels in the image, while KAN convolution can apply a learnable nonlinear activation function to each element and then sum it. Learnability means that it provides more expressiveness for the activation function. Compared with traditional convolution, KAN convolution can effectively reduce the number of model parameters while maintaining or improving feature extraction capabilities, thereby improving training efficiency and inference speed. These advantages make the application of KAN convolution in U-Net particularly suitable for tasks that require high-precision feature extraction, such as image segmentation.

[0055] 2-3) Fusion: The feature map X1 obtained by wavelet convolution and the feature map X2 obtained by KAN convolution are fused to generate a new feature Y. The attention mechanism is further introduced. The attention mechanism emphasizes the key feature area and suppresses redundant information by weighting. It can help the model more effectively integrate features from different convolutional layers, ensure that key information is retained and strengthened, and promote more accurate segmentation results. CBAM is an attention mechanism used to improve the performance of convolutional neural networks. It enhances the expressiveness of feature maps through two key steps: channel attention and spatial attention. In the channel attention module, CBAM uses global average pooling and global maximum pooling to extract global features of each channel, and generates channel weights through a shared multi-layer perceptron (MLP) to emphasize feature channels that are important to the task. Next, the spatial attention module processes the feature map after channel weighting, generates a spatial attention map through convolution operation, and further highlights the key spatial area. Through this mechanism, CBAM effectively focuses on important features and reduces the impact of redundant information;

[0056] 2-4) Generate final features: The fused feature Y is processed by the attention mechanism CBAM to generate the final feature map Y1;

[0057] The characteristic changes are shown as follows Figure 4 As shown in the figure, the features of wavelet transform convolution are fused with the features of KAN convolution, and finally the feature map is weighted by the attention mechanism to highlight important features and suppress irrelevant features, thereby enhancing the model's sensitivity to important information and further improving the feature expression ability. The WKA module can effectively extract global information while focusing on important features, reducing redundant information images, and fully capturing the edge features of multifidus muscle MRI images. This WKA-Unet combination strategy enables the model to have stronger capabilities when facing complex image segmentation tasks;

[0058] 3) Testing: In order to evaluate the segmentation performance of different models, a variety of evaluation indicators are used to evaluate the models, including the Dice coefficient, the similarity coefficient Jaccard and the Hausdorff distance HD. The Dice coefficient is used to quantify the similarity between the predicted segmentation area and the actual label area. The Dice coefficient value ranges from 0 to 1. The closer the value is to 1, the more consistent the segmentation result is with the actual situation. The Jaccard index is the intersection over union ratio IoU, which is used to measure the degree of overlap between samples. The Jaccard index value is between 0 and 1. The larger the value, the more obvious the overlap. The Hausdorff distance HD reflects the similarity between point sets. The maximum distance between the actual annotation and the predicted segmentation boundary is calculated to evaluate the segmentation accuracy. It is particularly sensitive to the accuracy of the boundary. The smaller the value, the higher the segmentation accuracy. The calculation method of the evaluation indicators is as follows:

[0059] The calculation formula of DSC is:

[0060]

[0061] The calculation formula of JAC is:

[0062]

[0063] HD calculation formula:

[0064] HD(p,g)=max(h(p,g),h(g,p))

[0065]

[0066]

[0067] Where p is the pixel set of the predicted segmentation, and g is the pixel set of the Ground Truth.

[0068] In this example, the experiment selected Ubuntu 16.04 as the operating system environment, equipped with an Nvidia RTX3090 GPU with 8GB of video memory, and used the PyTorch 1.1 deep learning framework and Python. 3.6 programming language, built and implemented a neural network model for training and testing, compared the final segmentation results with the true label values, and obtained three parameter values, Dice coefficient DSC, JAC and HD. DSC measures the degree of overlap. The obtained results are compared with the results of segmentation by other models. If they are better than other models, they meet expectations. If they do not meet expectations, adjust the parameters and retrain the unet network until the expected results are achieved. The output model and the segmentation results of the verification experiment are shown in Table 1. Table 1 lists the quantitative results of DSC, JAC and HD. The FCN and U-Net models are compared and analyzed. Compared with FCN, this method improves by 20.93% on DSC, 35.48% on JAC, and decreases by 15.06% on HD indicators; compared with U-Net, this method improves by 14.76% on DSC, 15.86% on JAC, and decreases by 16.77% on HD indicators. Experimental results show that the segmentation accuracy of this method is better than that of general methods. Compared with other existing segmentation algorithms, this method has better semantic segmentation performance in details such as object edges, and has improved performance in segmentation evaluation indicators DSC and JAC.

[0069] Table 1

[0070]

Claims

1. A WKA-Unet-based MRI image segmentation method, characterized in that: The steps include: 1) Data preprocessing: Collect image data sets, and preprocess the data sets including abnormal image processing, Gaussian noise, size scaling, contrast adjustment, and data enhancement. Specifically: 1-1) Abnormal image processing: remove abnormal images that are not suitable for training. Abnormal images include: images with severe imaging artifacts, images with low resolution and low quality, images that do not contain the complete target area, and images with severe lesions; 1-2) Gaussian noise: Add Gaussian noise to the image to simulate MRI images in a noisy environment. Generate a random number that conforms to the Gaussian distribution through a random number generator, add the generated random number to each pixel of the image, and use the numpy.random.normal() function in Numpy to generate noise that satisfies the Gaussian distribution and superimpose it on the image pixel value; 1-3) Resizing: Use the resize() function of the image processing library or programming language to adjust the image size. When the nearest neighbor interpolation method is used, for each target pixel, its value will be set to the value of the nearest input pixel. Specifically, adjust the image size and scale the input image to the specified target size. Assign pixel values. For each target pixel, determine the input pixel closest to it and assign the value of the input pixel to the target pixel. The image is finally cropped or scaled to 256×256 pixels. 1-4) Contrast adjustment: Use image contrast adjustment, namely histogram equalization and contrast stretching, specifically: perform histogram equalization on the image to enhance the contrast of the image; use OpenCV's cv2.equalizeHist() function to equalize the grayscale image; 1-5) Data enhancement: Geometric transformation and flipping are used to expand the data set. Geometric transformation rotates the MRI image 30 degrees clockwise through rotation operation, and flipping uses up and down horizontal mirroring operation to generate a mirror image; 1-6) Classification: Expand the image dataset after data preprocessing in steps 1-1) to 1-5) to 1,000 images, and divide them into a training set and a test set in a ratio of 7:3; 2) Constructing WKA-Unet network: WKA-Unet is improved based on Unet network, and the WKA module is introduced. The WKA module is embedded in each layer of Unet network. Unet network is a convolutional neural network architecture for biomedical image segmentation. The basic structure includes: Encoder: A series of convolutional pooling layers are used to capture the contextual information of the image, and convolutional pooling layers are used for downsampling operations; Decoder: A series of upsampling and convolutional layers are used to recover the fine details of the segmentation; Skip connection: copy the feature map from the contraction path to the corresponding expansion path to retain more detail information; The WKA module combines wavelet transform convolution, KAN convolution and CBAM attention mechanism, including: 2-1) Wavelet transform convolution: The input feature map X is first processed by wavelet transform convolution to obtain feature map X1; 2-2) KAN convolution: The input feature map X is also processed by KAN convolution to obtain feature map X2; 2-3) Fusion: The feature map X1 obtained by wavelet convolution and the feature map X2 obtained by KAN convolution are fused to generate a new feature Y; 2-4) Generate final features: The fused feature Y is processed by the attention mechanism CBAM to generate the final feature map Y1; 3) Testing: A variety of evaluation indicators are used to evaluate the model, including the Dice coefficient, the similarity coefficient Jaccard and the Hausdorff distance HD. The Dice coefficient is used to quantify the similarity between the predicted segmentation area and the actual label area. The Dice coefficient value ranges from 0 to 1. The closer the value is to 1, the more consistent the segmentation result is with the actual situation. The Jaccard index is the intersection over union ratio IoU, which is used to measure the overlap between samples. The Jaccard index value is between 0 and 1. The larger the value, the more obvious the overlap. The Hausdorff distance HD reflects the similarity between point sets. The maximum distance between the actual annotation and the predicted segmentation boundary is calculated to evaluate the segmentation accuracy. It is particularly sensitive to the accuracy of the boundary. The smaller the value, the higher the segmentation accuracy. The calculation method of the evaluation indicators is as follows: The calculation formula of DSC is: The calculation formula of JAC is: HD calculation formula: HD(p,g)=max(h(p,g),h(g,p)) Where p is the pixel set of the predicted segmentation, and g is the pixel set of the Ground Truth.

Images