A pancreas segmentation method based on WEUnet network

Abstract

This invention discloses a pancreas segmentation method based on the WEUnet network, belonging to the field of pancreas segmentation technology. It solves the problem of unclear edges in pancreas segmentation. The method includes the following steps: S1, acquiring an abdominal CT image and inputting it into the WEUnet network; S2, performing five consecutive convolutional encoding processes on the abdominal CT image through the encoding module to obtain the encoded visual feature maps output by each layer of the encoding module; S3, inputting the bottom-level encoded visual feature map into the decoding module to obtain the bottom-level decoded prediction feature map; S4, performing four consecutive convolutional decoding processes on the bottom-level decoded prediction feature map through the decoding module to obtain the top-level decoded prediction feature map; S5, obtaining the pancreas image based on the top-level decoded prediction feature map. This network effectively solves the problems of pancreas deformation, unclear edges, and class imbalance, and further improves the segmentation index.

Description

Technical Field

[0001] This invention belongs to the field of pancreatic segmentation technology, specifically relating to a pancreatic segmentation method based on the WEUnet network. Background Technology

[0002] In recent years, with the rapid development of deep learning and neural networks, abdominal organs such as the liver, lungs, and kidneys can be effectively segmented using artificial intelligence systems. However, the segmentation accuracy of the pancreas is still not high enough to be applied in clinical practice. The main difficulties stem from the following aspects:

[0003] 1) The pancreas occupies only a small portion of the entire CT image. The significant size imbalance between the target and background makes it easy for the network to focus on non-target background areas, leading to overfitting of the background region. 2) The pancreas has a highly variable shape. Its irregular shape and susceptibility to deformation result in significant differences in the shape, size, and location of the abdominal pancreas among different patients. 3) The pancreas's boundaries are unclear. Due to the similar density of the pancreas to surrounding tissues (duodenum, small intestine, etc.), the contrast between the pancreas and surrounding tissues is weak in CT images, resulting in boundary perturbation and poor boundary definition.

[0004] Traditional machine learning image segmentation algorithms struggle to accurately segment the blurred edges of the pancreas to address these issues. Currently, the main techniques for medical image segmentation are based on data-driven feature learning. One commonly used neural network architecture is the Convolutional Neural Network (CNN) U-Net. It consists of three parts: an encoder, a decoder, and skip connections. The encoder extracts high-level semantic features from the image while progressively reducing the spatial resolution. The decoder progressively restores the spatial resolution of the feature maps while preserving the semantic information extracted by the encoder. Skip connections directly pass features from corresponding layers in the encoder to the decoder, fusing low-level and high-level features to compensate for information loss.

[0005] While the combination of skip connections and upsampling works well, it still falls short in achieving optimal model performance. To address this issue, researchers have employed various strategies. Hu and Wang et al. have delved into spatial attention mechanisms in nonlocal neural networks, enabling neural networks to grasp long-range dependencies and complex spatial structures in images. Li et al. outlined a multi-scale attention-intensive residual network that enhances the model's ability to accurately locate subtle differences in the pancreas by merging dense residual blocks and multi-scale convolutional kernels. Yan et al. proposed an architecture that merges 2D and 3D convolutional layers, reducing computational requirements while preserving spatial information. Chen et al. introduced fuzzy logic into the skip connections of U-Net, improving the model's sensitivity to small, variable structures. Wang et al. added attention-guided control mechanisms to skip connections, enhancing the model's ability to recognize minute structures. Although attention mechanisms are advantageous in handling multi-scale objects, they remain limited by kernel size and face challenges in capturing features of elongated organs like the pancreas. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a pancreas segmentation method based on the WEUnet network, which solves the problem of unclear edges in pancreas segmentation.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a pancreas segmentation method based on the WEUnet network, comprising the following steps:

[0008] S1. Acquire abdominal CT images and input them into the WEUnet network;

[0009] S2. The abdominal CT image is subjected to five consecutive convolutional coding processes through the coding module to obtain the coded visual feature map output by each layer coding module.

[0010] S3. Input the bottom-level encoded visual feature map into the decoding module to obtain the bottom-level decoded prediction feature map;

[0011] S4. The decoding module performs four consecutive convolutional decoding processes on the bottom layer's decoding prediction feature map to obtain the top layer's decoding prediction feature map. During each decoding process, the edge feature attention module with skip connections calculates an intermediate feature map based on the current layer's encoded visual feature map, the next layer's decoding prediction feature map, and the current layer's high-frequency feature map. The intermediate feature map is then input into the decoding module, and the generated result is concatenated onto the generated current layer's decoding prediction feature map.

[0012] S5. Obtain the pancreas image based on the top-level decoded and predicted feature map.

[0013] Further: In S2, the encoding module includes a wavelet convolution submodule, a depth convolution submodule, and a pointwise convolution submodule connected in sequence;

[0014] The output of the wavelet convolution submodule is connected to the input of the first convolution submodule and the input of the first normalization layer, respectively. The outputs of the pointwise convolution submodule, the first convolution submodule, and the first normalization layer are all connected to the input of the first element-wise addition. The output of the first element-wise addition is connected to the inputs of the second convolution submodule and the second normalization layer through the first ReLU activation layer, respectively. The outputs of the second convolution submodule and the second normalization layer are both connected to the input of the second element-wise addition. The output of the second element-wise addition is connected to the input of the second ReLU activation layer.

[0015] The wavelet convolution submodule includes wavelet convolution, a third normalization layer, and a third ReLU activation layer connected in sequence.

[0016] The deep convolutional submodule and the pointwise convolutional submodule have the same structure, both including an interconnected first 3×3 convolution and a fourth normalization layer;

[0017] The first and second convolutional submodules have the same structure, both including interconnected first 1×1 convolutions and fifth normalization layers.

[0018] Further: In S2, the method for outputting the encoded visual feature map by the encoding module is specifically as follows:

[0019] S21. Obtain the first input feature map of the input encoding module, and input the first input feature map into the wavelet convolution sub-module to obtain the first feature map;

[0020] S22. The first feature map is sequentially input into the depth convolution submodule and the pointwise convolution submodule, and the residual branch is used to input the first feature map into the first convolution submodule and the first normalization layer respectively. The outputs of the pointwise convolution submodule, the first convolution submodule and the first normalization layer are added together by the first element to obtain the second feature map.

[0021] S23. The second feature map is passed through the first ReLU activation layer and the second convolutional sub-module in sequence, and the second feature map is input into the second normalization layer using the residual branch. The outputs of the second convolutional sub-module and the second normalization layer are added together by the second element to obtain the third feature map.

[0022] S24. Input the third feature map into the second ReLU activation layer to obtain the encoded visual feature map.

[0023] Further: S21 includes the following sub-steps:

[0024] S211. Generate a first low-frequency component and a first high-frequency component from the first input feature map through wavelet transform;

[0025] S212. Generate a second low-frequency component and a second high-frequency component from the first low-frequency component through wavelet transform.

[0026] S213. Perform convolution and transpose convolution on the second low-frequency component and the second high-frequency component in sequence, and take all the processed components as the convolution result of the second wavelet transform.

[0027] S214. Add the convolution result of the second wavelet transform to the first low-frequency component and the first high-frequency component. Perform convolution and transpose convolution on the result after addition in sequence. Use all processed components as the convolution result of the first wavelet transform.

[0028] S215. Add the convolution result of the first wavelet transform to the first input feature map, and then perform convolution processing on the result to obtain the first feature map.

[0029] Further: In S3, the decoding module includes a third convolution submodule, a fourth ReLU activation layer, and a fourth convolution submodule connected in sequence. The input of the third convolution submodule is also connected to the input of the fifth convolution submodule. The outputs of the fourth convolution submodule and the fifth convolution submodule are both connected to the input of the third element-adding layer. The output of the third element-adding layer is connected in sequence to the fifth ReLU activation layer and the upsampling operation layer.

[0030] The third and fourth convolutional submodules have the same structure, both including interconnected second 3×3 convolutions and a sixth normalization layer.

[0031] The fifth convolutional submodule consists of interconnected second 1×1 convolutions and a seventh normalized layer.

[0032] Further: In step S3, the method for processing the second input feature map through the decoding module is specifically as follows:

[0033] S31. Obtain the second input feature map of the input decoding module, and input the second input feature map into the third convolutional submodule, the fourth ReLU activation layer and the fourth convolutional module;

[0034] S32. The second input feature map is input into the fifth convolutional submodule using a residual branch;

[0035] S33. The outputs of the fourth and fifth convolutional sub-modules are added together by the third element to obtain the fourth feature map. The fourth feature map is then input into the fifth ReLU activation layer and the upsampling operation layer in sequence to obtain the result of the second input feature map processed by the decoding module.

[0036] Furthermore: In S4, the edge feature attention module is provided with a convolutional attention submodule, which includes a spatial attention submodule interconnected with the channel attention submodules;

[0037] The channel attention submodule is configured with a first convolution with a 1×1×C kernel, and the spatial attention submodule is configured with a second convolution with an H×W×1 kernel.

[0038] Furthermore: In step S4, the high-frequency feature map is obtained by processing the abdominal CT image using DWT;

[0039] The method for calculating intermediate feature maps during each decoding process includes the following steps:

[0040] S41. Obtain the encoded visual feature map of the current layer, the high-frequency feature map of the current layer, and the decoding prediction feature map of the next layer. Multiply the high-frequency feature map of the current layer and the decoding prediction feature map of the next layer element by element with the encoded visual feature map of the current layer. Concatenate the generated feature map with the encoded visual feature map of the current layer to generate a fused feature map.

[0041] S42. Input the fused feature map into the channel attention submodule, and use the first branch to process the fused feature map by max pooling and multilayer perceptron in sequence. After max pooling and multilayer perceptron processing, optimize it by the first convolution to generate the first attention feature map.

[0042] S43. The second branch is used to process the fused feature map by average pooling and multilayer perceptron in sequence, and after average pooling and multilayer perceptron processing, it is optimized by the first convolution to generate the second attention feature map.

[0043] S44. Add the first attention feature map and the second attention feature map, process the result of the addition with the Sigmoid activation function and the first convolution optimization, and multiply the optimization result with the fused feature map to generate the third attention feature map.

[0044] S45. Input the third attention feature map into the spatial attention submodule, and use the third branch to perform max pooling and second convolution optimization on the third attention feature map to generate the fourth attention feature map; use the fourth branch to perform average pooling and second convolution optimization on the third attention feature map to generate the fifth attention feature map.

[0045] S46. Fuse the fourth attention feature map and the fifth attention feature map. The fusion result is then subjected to second convolution optimization, sigmoid activation function processing and second convolution optimization in sequence. The optimization result is multiplied with the third attention feature map to generate an intermediate feature map.

[0046] The beneficial effects of this invention are as follows:

[0047] (1) This invention provides a pancreas segmentation method based on the WEUnet network. When dealing with complex pancreas segmentation tasks, this network effectively solves the problems of pancreas deformation, unclear edges, and class imbalance, and further improves the segmentation index. This invention proposes a wavelet convolutional submodule using Dobesie wavelet transform. By fusing low-frequency and high-frequency information, the network focuses more on the edges of the pancreas and can better integrate local and global features, making the network expression clearer.

[0048] (2) The edge feature attention module proposed in this invention makes good use of the additional high-frequency information, which solves the problem of boundary ambiguity to a certain extent. In addition, the transfer learning strategy adopted reduces training costs and improves generalization ability. The method of this invention was evaluated on the publicly available NIH dataset and MSD dataset. Experimental results show that the proposed WEUnet network outperforms other methods in segmentation results on both datasets, proving that the proposed network is effective.

[0049] (3) This invention proposes a novel neural network called WEUNet, which achieves performance comparable to mainstream methods while reducing the number of parameters and lowering training costs. It achieves an average DSC score of 87.61±2.34% on the NIH dataset and an average DSC score of 87.29±2.52% on the MSD dataset. Attached Figure Description

[0050] Figure 1 This is a flowchart of a pancreas segmentation method based on the WEUnet network according to the present invention.

[0051] Figure 2 This is a diagram showing the overall structure of the WEUnet network.

[0052] Figure 3 This is a schematic diagram of the encoding module.

[0053] Figure 4 This is a schematic diagram of a wavelet convolutional submodule.

[0054] Figure 5 This is a schematic diagram of the decoding module.

[0055] Figure 6 This is a schematic diagram of the edge feature attention module.

[0056] Figure 7 This is a boxplot representation of four-fold cross-validation on the NIH dataset.

[0057] Figure 8 This is the segmentation result from the NIH dataset. Detailed Implementation

[0058] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0059] like Figure 1 and Figure 2 As shown, in one embodiment of the present invention, a pancreas segmentation method based on the WEUnet network includes the following steps:

[0060] S1. Acquire abdominal CT images and input them into the WEUnet network;

[0061] S2. The abdominal CT image is subjected to five consecutive convolutional coding processes through the coding module to obtain the coded visual feature map output by each layer coding module.

[0062] S3. Input the bottom-level encoded visual feature map into the decoding module to obtain the bottom-level decoded prediction feature map;

[0063] S4. The decoding module performs four consecutive convolutional decoding processes on the bottom layer's decoding prediction feature map to obtain the top layer's decoding prediction feature map. During each decoding process, the edge feature attention module with skip connections calculates an intermediate feature map based on the current layer's encoded visual feature map, the next layer's decoding prediction feature map, and the current layer's high-frequency feature map. The intermediate feature map is then input into the decoding module, and the generated result is concatenated onto the generated current layer's decoding prediction feature map.

[0064] S5. Obtain the pancreas image based on the top-level decoded and predicted feature map.

[0065] In this embodiment, the overall structure diagram is as follows: Figure 2 As shown, it adopts a U-shaped jump connection structure, with five encoding modules and five decoding modules connected in series in the encoding path and decoding path, respectively.

[0066] The encoding module is based on the MobileOne architecture, leveraging its lightweight and efficient characteristics, and introduces a wavelet convolution submodule to enhance the extraction capability of specific features. Specifically, in the main path, the input feature map first passes through a wavelet convolution WTconv, followed by normalized Norm and ReLU activation. Subsequently, the feature map passes through depthwise convolution and pointwise convolution in sequence, with the specific structure consistent with MobileOne. This allows for the learning of richer features during downsampling by replacing pooling layers with depthwise separable convolutions.

[0067] In the decoding module, the feature map is processed by two consecutive 3×3 convolutions, followed by normalization and ReLU activation. To prevent network degradation, a residual branch is introduced, thereby further improving the feature restoration and fusion capabilities.

[0068] Traditional CNN architectures often lose crucial information when downsampling to deeper levels, information that is essential for medical segmentation. Furthermore, to enhance the learning ability of small objects and edge features, an Edge Feature Attention Module (SEA) with residual properties is designed between sibling encoder-decoder blocks in skip connections. This module operates between upsampling and downsampling paths at each resolution level, receiving three inputs: the output of each layer in the downsampling path, the output of the next lower layer in the upsampling path, and additional extracted edge features. The Edge Feature Attention Module uses these inputs to build a coherent bridge between the contraction and expansion paths. This design not only effectively preserves and emphasizes edge features but also enhances the representation of small objects, thereby significantly improving the network's ability to capture both detailed and global features in medical segmentation tasks.

[0069] like Figure 3 As shown, in S2, the encoding module includes a wavelet convolution submodule, a depth convolution submodule, and a pointwise convolution submodule connected in sequence;

[0070] The output of the wavelet convolution submodule is connected to the input of the first convolution submodule and the input of the first normalization layer, respectively. The outputs of the pointwise convolution submodule, the first convolution submodule, and the first normalization layer are all connected to the input of the first element-wise addition. The output of the first element-wise addition is connected to the inputs of the second convolution submodule and the second normalization layer through the first ReLU activation layer, respectively. The outputs of the second convolution submodule and the second normalization layer are both connected to the input of the second element-wise addition. The output of the second element-wise addition is connected to the input of the second ReLU activation layer.

[0071] like Figure 4 As shown, the wavelet convolution submodule includes wavelet convolution, a third normalization layer, and a third ReLU activation layer connected in sequence;

[0072] The deep convolutional submodule and the pointwise convolutional submodule have the same structure, both including an interconnected first 3×3 convolution and a fourth normalization layer;

[0073] The first and second convolutional submodules have the same structure, both including interconnected first 1×1 convolutions and fifth normalization layers.

[0074] In this embodiment, traditional convolutional networks face significant challenges in capturing the unique features of the pancreas due to the blurred boundaries between it and adjacent organs. This is primarily because the pancreas has a complex anatomical location and a small size in the image, making it susceptible to background noise and interference from adjacent organs. To address these issues, this invention introduces wavelet convolution technology and improves it to suit the characteristics of small abdominal organs, thereby enhancing the network's ability to learn pancreatic features. Discrete Wavelet Transform (DWT) is a multi-resolution analysis tool that decomposes an input signal or image into low-frequency and high-frequency information. In image processing, DWT provides a way to decompose an image into different frequency components while preserving spatial information. This decomposition generates a set of sub-images with different details and resolutions. Specifically, the original image is decomposed into low-frequency and high-frequency components. The low-frequency component LL represents the overall structural information of the image, containing lower frequency content, equivalent to presenting the image at a lower spatial resolution. The high-frequency components include LH, HL, and HH: describing the details and textures in the image, where LH represents horizontal details, HL represents vertical details, and HH represents details in the diagonal direction.

[0075] In step S2, the method for outputting the encoded visual feature map by the encoding module is specifically as follows:

[0076] S21. Obtain the first input feature map of the input encoding module, and input the first input feature map into the wavelet convolution sub-module to obtain the first feature map;

[0077] S22. The first feature map is sequentially input into the depth convolution submodule and the pointwise convolution submodule, and the residual branch is used to input the first feature map into the first convolution submodule and the first normalization layer respectively. The outputs of the pointwise convolution submodule, the first convolution submodule and the first normalization layer are added together by the first element to obtain the second feature map.

[0078] S23. The second feature map is passed through the first ReLU activation layer and the second convolutional sub-module in sequence, and the second feature map is input into the second normalization layer using the residual branch. The outputs of the second convolutional sub-module and the second normalization layer are added together by the second element to obtain the third feature map.

[0079] S24. Input the third feature map into the second ReLU activation layer to obtain the encoded visual feature map.

[0080] S21 includes the following sub-steps:

[0081] S211. Generate a first low-frequency component and a first high-frequency component from the first input feature map through wavelet transform;

[0082] S212. Generate a second low-frequency component and a second high-frequency component from the first low-frequency component through wavelet transform.

[0083] S213. Perform convolution and transpose convolution on the second low-frequency component and the second high-frequency component in sequence, and take all the processed components as the convolution result of the second wavelet transform.

[0084] S214. Add the convolution result of the second wavelet transform to the first low-frequency component and the first high-frequency component. Perform convolution and transpose convolution on the result after addition in sequence. Use all processed components as the convolution result of the first wavelet transform.

[0085] S215. Add the convolution result of the first wavelet transform to the first input feature map, and then perform convolution processing on the result to obtain the first feature map.

[0086] In this embodiment, the first input feature map is input into the wavelet convolution submodule to obtain the first feature map. This method combines global information from low-frequency components with detailed features from high-frequency components, improving the model's ability to recognize blurred boundaries. Based on the first input feature map T, four sets of filters are used:

[0087]

[0088] In the formula, f LL It is a low-pass filter, f LH f HL f HH It is a high-pass filter.

[0089] Then, the above filter is used to convolve each channel to output low-frequency and high-frequency components.

[0090] [Z LL Z LH Z HL Z HH ] = Conv([f LL ,f LH ,f HL ,f HH ],T)

[0091] In the formula, Z LL It is the low-frequency component of T, Z LH Z HL Z HH These are its high-frequency components in the horizontal, vertical, and diagonal directions, respectively, and Conv(·) is the convolution process;

[0092] Finally, the processed components are restored by transposed convolution to generate the wavelet transform convolution result Z.

[0093] Z = Convtransposed([f LL ,f LH ,fHL ,f HH ],Z c )

[0094] Z c =Concat(Z) LL Z LH Z HL Z HH )

[0095] In the formula, Convtransposed(·) represents the transposed convolution process, and Z c This represents the result of connecting low-frequency and high-frequency components; Concat is the connection operation.

[0096] like Figure 5 As shown, in S3, the decoding module includes a third convolution submodule, a fourth ReLU activation layer, and a fourth convolution submodule connected in sequence. The input of the third convolution submodule is also connected to the input of the fifth convolution submodule. The outputs of the fourth convolution submodule and the fifth convolution submodule are both connected to the input of the third element-adding layer. The output of the third element-adding layer is connected in sequence to the fifth ReLU activation layer and the upsampling operation layer.

[0097] The third and fourth convolutional submodules have the same structure, both including interconnected second 3×3 convolutions and a sixth normalization layer.

[0098] The fifth convolutional submodule consists of interconnected second 1×1 convolutions and a seventh normalized layer.

[0099] In step S3, the method for processing the second input feature map through the decoding module is as follows:

[0100] S31. Obtain the second input feature map of the input decoding module, and input the second input feature map into the third convolutional submodule, the fourth ReLU activation layer and the fourth convolutional module;

[0101] S32. The second input feature map is input into the fifth convolutional submodule using a residual branch;

[0102] S33. The outputs of the fourth and fifth convolutional sub-modules are added together by the third element to obtain the fourth feature map. The fourth feature map is then input into the fifth ReLU activation layer and the upsampling operation layer in sequence to obtain the result of the second input feature map processed by the decoding module.

[0103] like Figure 6 As shown, in S4, the edge feature attention module is provided with a convolutional attention submodule, which includes a spatial attention submodule interconnected with the channel attention submodules;

[0104] The channel attention submodule is configured with a first convolution with a 1×1×C kernel, and the spatial attention submodule is configured with a second convolution with an H×W×1 kernel.

[0105] In this embodiment, the primary goal of the edge feature attention module is to stably preserve high-frequency information across multiple scales, thereby effectively addressing the boundary ambiguity problem in pancreas segmentation. Furthermore, the edge feature attention module plays a crucial role in bridging the semantic gap between low-level features extracted by the encoder and high-level features generated by the decoder before their fusion.

[0106] In S4, the high-frequency feature map is obtained by processing abdominal CT images using DWT.

[0107] The method for calculating intermediate feature maps during each decoding process includes the following steps:

[0108] S41. Obtain the encoded visual feature map of the current layer, the high-frequency feature map of the current layer, and the decoded prediction feature map of the next layer. Multiply the high-frequency feature map of the current layer and the decoded prediction feature map of the next layer element-wise with the encoded visual feature map of the current layer. Concatenate the generated feature map with the encoded visual feature map of the current layer to generate a fused feature map f. i a ;

[0109]

[0110] In the formula, f i E To obtain the encoded visual feature map of the current layer, f i DWT This is the high-frequency feature map of the current layer. This is the decoded and predicted feature map for the next layer, where i is the current layer ordinal number. [] represents element-wise multiplication, and [] represents feature concatenation;

[0111] After DWT processing, multiple high-frequency components are obtained; in this embodiment, the vertical high-frequency component is used. The high-frequency component obtained in the i-th layer needs to be obtained by filtering the features of the (i-1)-th layer and then downsampling by a factor of 2. However, this process may reduce the high-frequency information intensity of the i-th layer. To preserve the high-frequency information f in each layer... i DWT The present invention is from The high-frequency features of each layer are derived to most effectively preserve high-frequency information.

[0112] f i DWT =(down(f l )) i if i≥1

[0113] In the formula, down represents double downsampling, (down(f l )) i f l After i downsampling operations, f l This is a high-frequency feature map, which is obtained by processing abdominal CT images using DWT.

[0114] S42. Input the fused feature map into the channel attention submodule, and use the first branch to process the fused feature map by max pooling and multilayer perceptron in sequence. After max pooling and multilayer perceptron processing, optimize it by the first convolution to generate the first attention feature map.

[0115] S43. The second branch is used to process the fused feature map by average pooling and multilayer perceptron in sequence, and after average pooling and multilayer perceptron processing, it is optimized by the first convolution to generate the second attention feature map.

[0116] S44. Add the first attention feature map and the second attention feature map, process the result of the addition with the Sigmoid activation function and the first convolution optimization, and multiply the optimization result with the fused feature map to generate the third attention feature map.

[0117] S45. Input the third attention feature map into the spatial attention submodule, and use the third branch to perform max pooling and second convolution optimization on the third attention feature map to generate the fourth attention feature map; use the fourth branch to perform average pooling and second convolution optimization on the third attention feature map to generate the fifth attention feature map.

[0118] S46. Fuse the fourth attention feature map and the fifth attention feature map. The fusion result is then subjected to second convolution optimization, sigmoid activation function processing and second convolution optimization in sequence. The optimization result is multiplied with the third attention feature map to generate an intermediate feature map.

[0119] In this embodiment, to capture the feature correlation between the boundary region and the background region, the fused feature map is input into the Convolutional Attention Module (CBAM) for recalibration. CBAM comprises two consecutive modules: a Channel Attention module focusing on the channel dimension, using a 1×1×C convolutional kernel to optimize the attention feature map; and a Spatial Attention module focusing on the spatial dimension, using an H×W×1 convolutional kernel to optimize the attention feature map. Through this process, optimized decoded features are obtained.

[0120] To evaluate the advantages of the model of this invention from different perspectives, it is first compared with networks that perform well on the NIH dataset. Then, the segmentation results of the proposed WEUnet network are visualized.

[0121] Table 1 shows the comparison results of the proposed WEUnet network with existing networks on the NIH dataset using four-fold cross-validation. The comparison results show that the proposed WEUnet network outperforms other methods, achieving a best DSC of 87.61%. A standard deviation of 2.34 indicates that the proposed method has high robustness across different cases on the NIH dataset. Furthermore, the recall rate of 89.92% demonstrates that the proposed model exhibits excellent pixel-level pancreas recognition capabilities. Moreover, compared to other models, the proposed model has fewer parameters and is more lightweight while maintaining a leading DSC performance.

[0122] Table 1. Pancreas segmentation results from the NIH dataset.

[0123]

[0124]

[0125] Table 2 provides a comparison of model parameters, showing the model size in more detail. The model of this invention performs exceptionally well due to its simple structure and smaller parameter count, with only 6.1M parameters. While maintaining leading DSC performance, the model of this invention has approximately four times fewer parameters than the best-performing model. This substantial improvement is attributed to transfer learning and fine-tuning techniques, which have consistently proven their effectiveness in the field of medical image segmentation.

[0126] Table 2 Comparison of Model Parameters (NIH Dataset)

[0127]

[0128] from Figure 7 As can be seen, the network of this invention achieves excellent performance on the NIH dataset, and the distribution of DSC values ​​in different sample sets is very similar, which indicates that the proposed network has high robustness and can mitigate the impact of sample variations.

[0129] Figure 8 The segmentation results of the WEUnet network on the NIH dataset are shown. The visualization of the segmentation results and the presentation of the DSC index demonstrate that the predictions of this invention are very close to reality, which means that the network of this invention can effectively capture the differences in pancreatic shape and size between individuals.

[0130] Table 3 shows the comparison results between the proposed WEUnet network and current methods that perform well on the MSD dataset. WEUnet achieves a Dice similarity coefficient of 87.29±2.52, leading in segmentation accuracy. Furthermore, WEUnet achieves precision of 86.37±2.15 and recall of 88.94±1.67, demonstrating high segmentation accuracy and low false negative rate. In summary, WEUnet maintains high accuracy while exhibiting strong robustness, fully demonstrating its effectiveness and reliability in segmentation tasks.

[0131] Table 3. Pancreas segmentation results from the MSD dataset.

[0132]

[0133] The beneficial effects of this invention are as follows: This invention provides a pancreas segmentation method based on the WEUnet network. When dealing with complex pancreas segmentation tasks, this network effectively solves the problems of pancreas deformation, unclear edges, and class imbalance, and further improves the segmentation index. This invention proposes a wavelet convolutional submodule using Dobesie wavelet transform. By fusing low-frequency and high-frequency information, the network focuses more on the edges of the pancreas and can effectively integrate local and global features, making the network representation clearer.

[0134] The edge feature attention module proposed in this invention effectively utilizes the additional high-frequency information, thus addressing the boundary ambiguity problem to some extent. Furthermore, the adopted transfer learning strategy reduces training costs and improves generalization ability. The method of this invention was evaluated on the publicly available NIH and MSD datasets. Experimental results show that the proposed WEUnet network outperforms other methods in segmentation results on both datasets, demonstrating the effectiveness of the proposed network.

[0135] This invention proposes a novel neural network called WEUNet, which achieves performance comparable to mainstream methods while reducing the number of parameters and lowering training costs. It achieves an average DSC score of 87.61±2.34% on the NIH dataset and an average DSC score of 87.29±2.52% on the MSD dataset.

[0136] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

Claims

1. A pancreas segmentation method based on WEUnet network, characterized in that, Includes the following steps: S1. Acquire abdominal CT images and input them into the WEUnet network; S2. The abdominal CT image is subjected to five consecutive convolutional coding processes through the coding module to obtain the coded visual feature map output by each layer coding module. S3. Input the bottom-level encoded visual feature map into the decoding module to obtain the bottom-level decoded prediction feature map; S4. The decoding module performs four consecutive convolutional decoding processes on the bottom layer's decoding prediction feature map to obtain the top layer's decoding prediction feature map. During each decoding process, the edge feature attention module with skip connections calculates an intermediate feature map based on the current layer's encoded visual feature map, the next layer's decoding prediction feature map, and the current layer's high-frequency feature map. The intermediate feature map is then input into the decoding module and concatenated onto the generated current layer's decoding prediction feature map. The high-frequency feature map was obtained by processing abdominal CT images using DWT. S5. Obtain the pancreas image based on the top-level decoded and predicted feature map.

2. The pancreas segmentation method based on WEUnet network according to claim 1, characterized in that, In S2, the encoding module includes a wavelet convolution submodule, a depth convolution submodule, and a pointwise convolution submodule connected in sequence; The output of the wavelet convolution submodule is connected to the input of the first convolution submodule and the input of the first normalization layer, respectively. The outputs of the pointwise convolution submodule, the first convolution submodule, and the first normalization layer are all connected to the input of the first element-wise addition. The output of the first element-wise addition is connected to the inputs of the second convolution submodule and the second normalization layer through the first ReLU activation layer, respectively. The outputs of the second convolution submodule and the second normalization layer are both connected to the input of the second element-wise addition. The output of the second element-wise addition is connected to the input of the second ReLU activation layer. The wavelet convolution submodule includes wavelet convolution, a third normalization layer, and a third ReLU activation layer connected in sequence. The deep convolutional submodule and the pointwise convolutional submodule have the same structure, both including an interconnected first 3×3 convolution and a fourth normalization layer; The first and second convolutional submodules have the same structure, both including interconnected first 1×1 convolutions and fifth normalization layers.

3. The pancreas segmentation method based on WEUnet network according to claim 2, characterized in that, In step S2, the method for outputting the encoded visual feature map by the encoding module is specifically as follows: S21. Obtain the first input feature map of the input encoding module, and input the first input feature map into the wavelet convolution sub-module to obtain the first feature map; S22. The first feature map is sequentially input into the depth convolution submodule and the pointwise convolution submodule, and the residual branch is used to input the first feature map into the first convolution submodule and the first normalization layer respectively. The outputs of the pointwise convolution submodule, the first convolution submodule and the first normalization layer are added together by the first element to obtain the second feature map. S23. The second feature map is passed through the first ReLU activation layer and the second convolutional sub-module in sequence, and the second feature map is input into the second normalization layer using the residual branch. The outputs of the second convolutional sub-module and the second normalization layer are added together by the second element to obtain the third feature map. S24. Input the third feature map into the second ReLU activation layer to obtain the encoded visual feature map.

4. The pancreas segmentation method based on WEUnet network according to claim 3, characterized in that, S21 includes the following sub-steps: S211. Generate a first low-frequency component and a first high-frequency component from the first input feature map through wavelet transform; S212. Generate a second low-frequency component and a second high-frequency component from the first low-frequency component through wavelet transform. S213. Perform convolution and transpose convolution on the second low-frequency component and the second high-frequency component in sequence, and take all the processed components as the convolution result of the second wavelet transform. S214. Add the convolution result of the second wavelet transform to the first low-frequency component and the first high-frequency component. Perform convolution and transpose convolution on the result after addition in sequence. Use all processed components as the convolution result of the first wavelet transform. S215. Add the convolution result of the first wavelet transform to the first input feature map, and then perform convolution processing on the result to obtain the first feature map.

5. The pancreas segmentation method based on WEUnet network according to claim 1, characterized in that, In S3, the decoding module includes a third convolution submodule, a fourth ReLU activation layer, and a fourth convolution submodule connected in sequence. The input of the third convolution submodule is also connected to the input of the fifth convolution submodule. The outputs of the fourth convolution submodule and the fifth convolution submodule are both connected to the input of the third element addition. The output of the third element addition is connected in sequence to the fifth ReLU activation layer and the upsampling operation layer. The third and fourth convolutional submodules have the same structure, both including interconnected second 3×3 convolutions and a sixth normalization layer. The fifth convolutional submodule consists of interconnected second 1×1 convolutions and a seventh normalized layer.

6. The pancreas segmentation method based on WEUnet network according to claim 5, characterized in that, In step S3, the method for processing the second input feature map through the decoding module is as follows: S31. Obtain the second input feature map of the input decoding module, and input the second input feature map into the third convolutional submodule, the fourth ReLU activation layer and the fourth convolutional module; S32. The second input feature map is input into the fifth convolutional submodule using a residual branch; S33. The outputs of the fourth and fifth convolutional sub-modules are added together by the third element to obtain the fourth feature map. The fourth feature map is then input into the fifth ReLU activation layer and the upsampling operation layer in sequence to obtain the result of the second input feature map processed by the decoding module.

7. The pancreas segmentation method based on WEUnet network according to claim 1, characterized in that, In S4, the edge feature attention module is provided with a convolutional attention submodule, which includes a spatial attention submodule that is interconnected with the channel attention submodules; The channel attention submodule is configured with 1×1× C The first convolution of the convolution kernel, the spatial attention submodule is set with H × W The second convolution with a ×1 convolution kernel.

8. The pancreas segmentation method based on WEUnet network according to claim 7, characterized in that, In step S4, the method for calculating the intermediate feature map during each decoding process includes the following sub-steps: S41. Obtain the encoded visual feature map of the current layer, the high-frequency feature map of the current layer, and the decoding prediction feature map of the next layer. Multiply the high-frequency feature map of the current layer and the decoding prediction feature map of the next layer element by element with the encoded visual feature map of the current layer. Concatenate the generated feature map with the encoded visual feature map of the current layer to generate a fused feature map. S42. Input the fused feature map into the channel attention submodule, and use the first branch to process the fused feature map by max pooling and multilayer perceptron in sequence. After max pooling and multilayer perceptron processing, optimize it by the first convolution to generate the first attention feature map. S43. The second branch is used to process the fused feature map by average pooling and multilayer perceptron in sequence, and after average pooling and multilayer perceptron processing, it is optimized by the first convolution to generate the second attention feature map. S44. Add the first attention feature map and the second attention feature map, process the result of the addition with the Sigmoid activation function and the first convolution optimization in sequence, and multiply the optimization result with the fused feature map to generate the third attention feature map. S45. Input the third attention feature map into the spatial attention submodule, and use the third branch to perform max pooling and second convolution optimization on the third attention feature map to generate the fourth attention feature map; use the fourth branch to perform average pooling and second convolution optimization on the third attention feature map to generate the fifth attention feature map. S46. Fuse the fourth attention feature map and the fifth attention feature map. The fusion result is then subjected to second convolution optimization, sigmoid activation function processing and second convolution optimization in sequence. The optimization result is multiplied with the third attention feature map to generate an intermediate feature map.

Images