3D medical image fuzzy boundary segmentation method and system based on multi-feature perception

Abstract

The invention relates to the technical field of 3D medical image segmentation, in particular to a 3D medical image fuzzy boundary segmentation method and system based on multi-feature perception, and the method comprises the following steps: carrying out the preprocessing of a to-be-segmented 3D medical image, and obtaining a low-dimensional feature sequence; a medical 3D form capture attention module is constructed to process the feature sequence, and form attention features are obtained; constructing a 3D image boundary perception segmentation network, and inputting the feature sequence into the 3D image boundary perception segmentation network for feature extraction to obtain boundary perception features; and constructing a 3D image boundary perception discriminator, splicing and fusing the morphological attention features and the boundary perception features to obtain a segmentation result, inputting the segmentation result into the 3D image boundary perception discriminator for evaluation, and feeding back and optimizing the 3D image boundary perception segmentation network according to the evaluation result. According to the invention, the fine heterogeneous texture features of the 3D medical image can be captured through the multi-feature perception attention module, the multi-level information of the 3D medical image is fully utilized, and the segmentation effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of 3D medical image segmentation, and in particular to a 3D medical image fuzzy boundary segmentation method and system based on multi-feature perception. Background Art

[0002] With the development of medical informatization technology, 3D image technology has played an important role in clinical practice. 3D medical image diagnosis is mainly used for early diagnosis, lesion localization, and treatment planning in clinical practice. It provides detailed three-dimensional structural information of organs, so that examiners can accurately observe the morphology, size, and location of organs, tumors, or other lesions, and then help doctors formulate personalized treatment plans. In addition, 3D images can also monitor the progression of diseases and help doctors improve the accuracy of clinical decisions.

[0003] However, 3D medical images may involve the three-dimensional spatial information of multiple different tissues or organs. Examiners need to analyze, extract, and process the details of each slice. Therefore, only by using 3D image-assisted segmentation technology can the efficiency of examination and analysis be effectively improved. With the development of artificial intelligence technology, deep learning has been gradually applied in the field of 3D medical image segmentation. Due to the possible blurring or complex shape of the 3D medical image boundary, general deep learning-based 3D medical image segmentation methods cannot effectively and accurately segment the image boundary. Traditional threshold methods and region growing methods rely on image gray differences, while fuzzy regions often exhibit features of gentle gradients and intensity overlaps, resulting in broken or over-segmented segmentation results. Moreover, active contour models are vulnerable to the initial position and are prone to falling into local optima in complex anatomical structures. The core challenges of medical image fuzzy boundaries are heterogeneity, deformability, and low signal-to-noise ratio.

[0004] Therefore, the present invention proposes a 3D medical image fuzzy boundary segmentation method and system based on multi-feature perception to solve the above problems. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention develops a 3D medical image fuzzy boundary segmentation method and system based on multi-feature perception. Through the multi-feature perception attention module, the present invention can capture the subtle heterogeneous texture features of 3D medical images, make full use of the multi-level information of 3D medical images, and improve the segmentation effect.

[0006] On the one hand, the technical solution for the present invention to solve the technical problem is a 3D medical image fuzzy boundary segmentation method based on multi-feature perception, including the following steps: S1. Preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; S2. Construct a medical 3D morphological capture attention module, input the low-dimensional feature sequence into the medical 3D morphological capture attention module to obtain morphological attention features; S3. Construct a 3D image boundary-aware segmentation network, which includes two multi-feature perception attention modules MFPA and one boundary-aware module BAM. Input the low-dimensional feature sequence into the 3D image boundary-aware segmentation network for feature extraction to obtain boundary-aware features; S4. Construct a 3D image boundary-aware discriminator, splice and fuse the morphological attention features and the boundary-aware features to obtain a segmentation result, then input the segmentation result into the 3D image boundary-aware discriminator for evaluation, and then optimize the 3D image boundary-aware segmentation network according to the evaluation result feedback.

[0007] S1 is specifically as follows: Obtain the 3D medical image to be segmented from the existing dataset. Any 3D medical image to be segmented is represented as , , where represents the height, represents the width, represents the depth, and represents the number of channels; Construct a 3D medical image preprocessing module, and input the 3D medical image to be segmented into the 3D medical image preprocessing module. Among them, through a sliding window of size slide in the image , and the image is divided into feature blocks, and each feature block is represented as , , . The calculation process of dividing the image into feature blocks is as follows: , where, represents the height of the sliding window, represents the width of the sliding window, represents the depth of the sliding window; Then, flatten the feature block into a one-dimensional vector , and the calculation formula is as follows: , where, represents the th feature block after being flattened into the corresponding one-dimensional vector , Represents a flattening operation; By linear transformation, the one-dimensional vector Projected into a low-dimensional feature space to obtain a fixed-length feature vector , , Represents the dimension after projection, and the calculation formula is as follows: , in, represents the linear projection matrix, , represents the bias vector, ; at last, A fixed-length feature vector Combine into a feature sequence , , .

[0008] S2 is as follows: Construct a medical 3D morphology capture attention module, which includes a normalization layer, a variable morphology capture attention module, and a convolution kernel size. The 3D convolution layer and convolution kernel size are The 3D convolution layer converts the low-dimensional feature sequence output by the 3D medical image preprocessing module Input to the medical 3D morphology capture attention module; Feature sequence First, it is normalized through the normalization layer to obtain the normalized features , and then input the normalized features into the variable form capture attention module, the normalized features After convolution kernel size is 3D convolutional layer and Activation function gets features , the calculation formula is as follows: ; Then, the features The convolution kernel size in the variable-shape capture attention module is 3D depthwise separable convolutional layer , the convolution kernel size is 3D dilated convolutional layer And the convolution kernel size is 3D convolutional layer , get the convolution feature , and then the convolution feature With features Splice and output features , and then the feature is input into the 3D convolutional layer with a convolutional kernel size of in the variable form capture attention module. The obtained feature is concatenated with the normalized feature to obtain the output feature of the variable form capture attention module. The calculation formula is as follows: , , , where represents the concatenation operation; Next, the output feature of the variable form capture attention module is concatenated with the low-dimensional feature sequence feature to obtain the enhanced feature after information supplementation. The enhanced feature passes through the 3D convolutional layer with a convolutional kernel size of and the 3D convolutional layer with a convolutional kernel size of to obtain the feature . The feature is concatenated with the enhanced feature to finally obtain the morphological attention feature output by the medical 3D form capture attention module. The calculation formula is as follows: , , .

[0009] S3 is specifically as follows: The two multi-feature perception attention modules MFPA are the first multi-feature perception attention module and the second multi-feature perception attention module respectively. Both multi-feature perception attention modules include a spatial attention feature capture module and a channel attention capture module inside. The input features pass through the spatial attention feature capture module and the channel attention capture module respectively to obtain the spatial attention feature and the channel attention feature, and then the two features are fused to obtain the output of the feature perception attention module; The boundary perception module concatenates and fuses the outputs of the two feature perception attention modules, constructs the boundary perception key feature map according to the intersection over union of the boundary region and the 3D true segmentation boundary image, and then calculates the boundary perception feature from the concatenated and fused features and the boundary perception key feature map.

[0010] The specific operations of the multi-feature perception attention module are as follows: The feature sequence Input into the first multi - feature perception attention module to obtain the first multi - feature perception feature , and then input the first multi - feature perception feature into the second multi - feature perception attention module to obtain the second multi - feature perception feature . The first multi - feature perception attention module and the second multi - feature perception attention module share the weights of the query and the key; The specific operation process in the first multi - feature perception attention module is as follows: (1) The morphological attention feature passes through the spatial attention feature capture module: Calculate the shared query , the shared key and the spatial value . The calculation formulas are as follows: , , , where , , represent the projection matrices of the query, the key, and the value respectively; Then, calculate the similarity between features through the dot - product of the shared query and the shared key to obtain the spatial attention feature map . Then multiply the spatial attention feature map by the spatial value to obtain the spatial attention feature . The calculation formulas are as follows: , , where represents the activation function, represents the transpose of the shared key , represents the dimension of the shared query and key; (2) The feature sequence passes through the channel attention feature capture module: , , , where and represent the query and key of the channel attention module, represents the channel value; By sharing the query and Share Key The dot product calculation is used to calculate the similarity between features and obtain the channel attention feature map , and then the channel attention feature map With channel value Multiply to get the channel attention feature , the calculation formula is as follows: , ; (3) Spatial attention features and channel attention features Add and fuse to get the fused features , the fused features Input into the convolution kernel size in sequence The convolutional layer and the convolution kernel size is The convolutional layer The final first multi-feature perception attention module features are obtained , the calculation formula is as follows: , ; The specific operation in the second multi-feature perception and attention module is the same as that in the first multi-feature perception and attention module. Input into the second multi-feature perception attention module to obtain the second multi-feature perception attention module feature .

[0011] The operations in the boundary perception module are as follows: The first multi-feature aware attention module features and the second multi-feature aware attention module features Splicing and fusion to generate perceptual fusion features ; Using Canny edge detection algorithm to segment 3D medical images Corresponding real 3D medical segmentation image Extract the boundary of the target object, obtain the target boundary contour of the 3D image, and randomly select candidate boundary key points, and obtain the key point set , Indicates The iterations are randomly selected A set of boundary points, , according to the key point set Construct the corresponding boundary area , represents the th key point, , , and respectively represent the x-axis, y-axis, and z-axis coordinates of the th key point randomly selected on the target boundary contour of the 3D image in the th iteration; Calculate the intersection over union (IoU) of the boundary region with the true 3D medical segmentation image , and the calculation formula is as follows: where , where represents the set of optimal boundary points obtained after calculating the IoU, which contains candidate boundary key points. Based on the set of optimal boundary points , construct the corresponding boundary-aware key feature map ; Fuse the key feature map with the perception fusion feature using channel weighting operations to obtain the boundary-aware feature , and the calculation process is as follows: , where represents element-wise addition, and represents channel-wise multiplication.

[0012] The process of calculating the segmentation result is specifically as follows: Concatenate the boundary-aware feature with the morphological attention feature in the channel dimension to obtain the fused feature . Adjust the number of channels of the fused feature using a convolutional layer with a kernel size of , and use activation function to output the final predicted segmentation 3D image , and the calculation formula is as follows: , , where represents the concatenation operation of features, and represents the convolutional operation with a kernel size of .

[0013] The 3D image boundary-aware discriminator is as follows: If the segmented 3D image finally predicted by the 3D image boundary-aware segmentation network The structural boundary of the retained target region is consistent with the real 3D medical segmentation image then the 3D image boundary-aware discriminator outputs a high score; conversely, if it fails to match the real segmentation image, it outputs a low score; among them, the discriminator module is to evaluate the quality of the predicted segmentation result output by the model, and the discriminator loss function is designed to enable the discriminator module to distinguish between high-score segmentation images and mis-segmented images. The calculation formula is as follows: Among them, represents the discriminator function, represents the score output by the discriminator, represents the concatenation operation, represents the convolutional layer with a convolutional kernel size of represents the fully connected layer, represents the Sigmoid activation function, represents the discriminator loss function, represents the logarithmic function with base 2; The 3D image boundary-aware segmentation network is optimized with the score output by the discriminator, and the 3D image boundary loss is introduced to adjust its own parameters according to the score. The calculation formula is as follows: Among them, represents the 3D image boundary loss, represents the logarithmic function with base 2.

[0014] On the other hand, the present invention also provides a 3D medical image fuzzy boundary segmentation system based on multi-feature perception, including modules for executing the processing instructions of each step in the 3D medical image fuzzy boundary segmentation method based on multi-feature perception; Image preprocessing module: Preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; Medical 3D morphology capture attention module: Input the low-dimensional feature sequence into the medical 3D morphology capture attention module to obtain morphological attention features; 3D image boundary-aware segmentation module: Includes two multi-feature perception attention modules MFPA and one boundary-aware module BAM. Input the morphological attention features into the 3D image boundary-aware segmentation module for feature extraction to obtain boundary-aware features; ​​​​​3D Image Boundary-Aware Discriminant Module: Concatenate and fuse the morphological attention features and boundary-aware features to obtain the segmentation result, then input the segmentation result into the 3D image boundary-aware discriminator for evaluation, and optimize the 3D image boundary-aware segmentation network according to the evaluation result feedback.

[0015] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: The present invention discloses a 3D medical image fuzzy boundary segmentation method and system based on multi-feature perception. The multi-feature perception attention module is used to capture the subtle heterogeneous texture features of 3D medical images. Combining with the boundary-aware module, the captured features are used to enhance the model's exploration of the fuzzy boundaries of objects, making full use of the multi-level information of 3D medical images, namely spatial-channel dual-path attention, and sharing weights, which can take into account both global and local features. It also actively learns high-confidence boundary key points through intersection over union screening and channel weighted fusion; the constructed medical 3D morphological capture attention module uses the deformable convolution mechanism to dynamically calibrate the modal feature heteroscedastic noise through the normalization layer and degree-scale feature concatenation, and effectively models the long-range dependence relationship and local structure deformation information between features in three-dimensional space. Through the combination of depthwise separable convolution and dilated convolution, the computational complexity is significantly reduced, and the model's expressiveness for complex organ boundaries and irregular morphologies is enhanced; at the same time, the present invention also introduces a 3D medical image boundary-aware discriminator to specifically optimize the fuzzy region, score the 3D image after the model segmentation, and feedback the scoring result to the model, effectively improving the segmentation effect of the 3D medical image fuzzy boundary. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation to the present invention.

[0017] Figure 1 It is a schematic flowchart of the method of the present invention.

[0018] Figure 2 It is a structural diagram of the system of the present invention.

[0019] Figure 3 It is a comparison diagram of the segmentation effects of the method of the present invention and the existing method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] In order to clearly illustrate the technical features of the present solution, the present invention will be elaborated in detail below through specific embodiments and in conjunction with its drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0021] Example 1 A 3D medical image fuzzy boundary segmentation method based on multi-feature perception, comprising the following steps: S1. Preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; S2. Construct a medical 3D shape capture attention module, input the low-dimensional feature sequence into the medical 3D shape capture attention module to obtain shape attention features; S3. Construct a 3D image boundary perception segmentation network, which includes two multi-feature perception attention modules MFPA and a boundary perception module BAM, input the low-dimensional feature sequence into the 3D image boundary perception segmentation network for feature extraction to obtain boundary perception features; S4. Construct a 3D image boundary perception discriminator, splice and fuse the shape attention features and the boundary perception features to obtain a segmentation result, then input the segmentation result into the 3D image boundary perception discriminator for evaluation, and then feedback and optimize the 3D image boundary perception segmentation network according to the evaluation result.

[0022] S1 is specifically as follows: Obtain the 3D medical image to be segmented from the existing dataset. Any 3D medical image to be segmented is represented as , , where represents the height, represents the width, represents the depth, represents the number of channels; Construct a 3D medical image preprocessing module, and input the 3D medical image to be segmented into the 3D medical image preprocessing module. Among them, through a sliding window of size slides in the image . The image is divided into feature blocks, and each feature block is represented as , , . The calculation process of dividing the image into feature blocks is as follows: , where, represents the height of the sliding window, represents the width of the sliding window, represents the depth of the sliding window; Then, flatten the feature block into a one-dimensional vector , the calculation formula is as follows: , Among them, represents the th feature block flattened into a corresponding one-dimensional vector , represents the flattening operation; After linearly transforming the one-dimensional vector and projecting it into a low-dimensional feature space, a feature vector with a fixed length is obtained, , represents the dimension after projection, and the calculation formula is as follows: , Among them, represents the linear projection matrix, , represents the bias vector, ; Finally, feature vectors with a fixed length are combined into a feature sequence , , .

[0023] S2 is specifically as follows: Construct a medical 3D shape capture attention module, which includes a normalization layer, a variable shape capture attention module, a 3D convolutional layer with a convolutional kernel size and a 3D convolutional layer with a convolutional kernel size of , and input the low-dimensional feature sequence output by the 3D medical image preprocessing module into the medical 3D shape capture attention module; The feature sequence is first normalized by the normalization layer to obtain a normalized feature , and then the normalized feature is input into the variable shape capture attention module. The normalized feature passes through a 3D convolutional layer with a convolutional kernel size of and the activation function to obtain a feature , and the calculation formula is as follows: ; Then, the feature successively passes through a 3D depthwise separable convolutional layer with a convolutional kernel size of and a 3D dilated convolutional layer with a convolutional kernel size of and a 3D convolutional layer with a convolutional kernel size of to obtain convolutional features . Then, the convolutional features are concatenated with the features to output features . Next, the features are input into a 3D convolutional layer with a convolutional kernel size of in the variable form capture attention module. The obtained features are concatenated with the normalized features to obtain the output features of the variable form capture attention module. The calculation formula is as follows: , , , where represents the concatenation operation; Next, the output features of the variable form capture attention module are concatenated with the low-dimensional feature sequence features to obtain the enhanced features after information supplementation. The enhanced features pass through a 3D convolutional layer with a convolutional kernel size of and a 3D convolutional layer with a convolutional kernel size of to obtain features . The features are concatenated with the enhanced features to finally obtain the morphological attention features output by the medical 3D form capture attention module. The calculation formula is as follows: , , .

[0024] S3 is specifically as follows: The two multi-feature perception attention modules MFPA are the first multi-feature perception attention module and the second multi-feature perception attention module respectively. Both multi-feature perception attention modules include a spatial attention feature capture module and a channel attention capture module inside. The input features pass through the spatial attention feature capture module and the channel attention capture module respectively to obtain spatial attention features and channel attention features, and then the two features are fused to obtain the output of the feature perception attention module; ​​​The boundary perception module splices and fuses the outputs of the two feature perception attention modules, constructs a boundary perception key feature map according to the intersection over union of the boundary region and the 3D true segmentation boundary image, and then calculates the boundary perception feature from the spliced and fused feature and the boundary perception key feature map.

[0025] The specific operations of the multi-feature perception attention module are as follows: Input the feature sequence into the first multi-feature perception attention module to obtain the first multi-feature perception feature , and then input the first multi-feature perception feature into the second multi-feature perception attention module to obtain the second multi-feature perception feature . The first multi-feature perception attention module and the second multi-feature perception attention module share the weights of the query and the key; The specific operation process in the first multi-feature perception attention module is as follows: (1) The morphological attention feature passes through the spatial attention feature capture module: Calculate the shared query , the shared key and the spatial value , and the calculation formulas are as follows: , , , Among them, , , respectively represent the projection matrices of the query, the key, and the value; Then, calculate the similarity between features through the dot product of the shared query and the shared key to obtain the spatial attention feature map , and then multiply the spatial attention feature map by the spatial value to obtain the spatial attention feature , and the calculation formulas are as follows: , , Among them, represents the activation function, represents the transpose of the shared key , represents the dimension of the shared query and key; (2) The feature sequence passes through the channel attention feature capture module: , , , Among them, and represent the query and key of the channel attention module, represents the channel value; By calculating the dot product of the shared query and the shared key , the similarity between features is calculated to obtain the channel attention feature map . Then, the channel attention feature map is multiplied by the channel value to obtain the channel attention feature . The calculation formula is as follows: , ; (3) Add and fuse the spatial attention feature and the channel attention feature to obtain the fused feature . The fused feature is sequentially input into the convolutional layer with a convolutional kernel size of and the convolutional layer with a convolutional kernel size of and the convolutional layer with a convolutional kernel size of to obtain the final first multi-feature perception attention module feature . The calculation formula is as follows: , ; The specific operations in the second multi-feature perception attention module are the same as those in the first multi-feature perception attention module. The first multi-feature perception attention module feature is input into the second multi-feature perception attention module to obtain the second multi-feature perception attention module feature .

[0026] The operations in the boundary perception module are as follows: Concatenate and fuse the first multi-feature perception attention module feature and the second multi-feature perception attention module feature to generate the perception fusion feature ; Use the Canny edge detection algorithm to obtain the corresponding real 3D medical segmentation image from the 3D medical image to be segmented ​​Extract the boundary of the target object to obtain the target boundary contour of the 3D image. Randomly select candidate boundary key points on the target boundary contour of the 3D image to obtain the key point set , indicating the th iteration of randomly selected boundary point sets, . According to the key point set , construct the corresponding boundary region , indicating the th key point, , , and respectively represent the x-axis, y-axis, and z-axis coordinates of the th randomly selected key point on the target boundary contour of the 3D image in the th iteration; Calculate the intersection over union of the boundary region and the real 3D medical segmentation image . The calculation formula is as follows: , where represents the best boundary point set obtained after calculating the intersection over union, containing candidate boundary key points. According to the best boundary point set , construct the corresponding boundary-aware key feature map ; Fuse the key feature map with the perception fusion feature using channel weighting operations to obtain the boundary-aware feature . The calculation process is as follows: , where represents element-wise addition, represents channel-wise multiplication.

[0027] The process of calculating the segmentation result is as follows: Concatenate the boundary-aware feature with the morphological attention feature in the channel dimension to obtain the fusion feature . Adjust the number of channels of the fused feature using a convolutional layer with a kernel size of , and use activation function to output the final predicted segmentation 3D image , the calculation formula is as follows: , , Among them, represents the splicing operation of features, represents the convolution operation with a convolution kernel size of .

[0028] The 3D image boundary-aware discriminator is as follows: If the segmented 3D image finally predicted by the 3D image boundary-aware segmentation network The structural boundary of the target region retained is consistent with the real 3D medical segmentation image , then the 3D image boundary-aware discriminator outputs a high score; otherwise, if it fails to match the real segmentation image, it outputs a low score; among them, the discriminator module is to evaluate the quality of the predicted segmentation result output by the model, and a discriminator loss function is designed so that the discriminator module can distinguish between high-score segmentation images and mis-segmented images. The calculation formula is as follows: , , , Among them, represents the discriminator function, represents the score output by the discriminator, represents the splicing operation, represents the convolution layer with a convolution kernel size of , represents the fully connected layer, represents the Sigmoid activation function, represents the discriminator loss function, represents the logarithm function with base 2; Use the score output by the discriminator to optimize the 3D image boundary-aware segmentation network, introduce the 3D image boundary loss, and adjust its own parameters according to the score. The calculation formula is as follows: , Among them, represents the 3D image boundary loss, represents the logarithm function with base 2.

[0029] Example 2 A 3D medical image fuzzy boundary segmentation system based on multi-feature perception includes a module for executing the processing instructions of each step in the 3D medical image fuzzy boundary segmentation method based on multi-feature perception; Image preprocessing module: Preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; Medical 3D Morphology Capture Attention Module: Input the low-dimensional feature sequence into the medical 3D morphology capture attention module to obtain morphological attention features; 3D Image Boundary-Aware Segmentation Module: It includes two Multi-Feature Perception Attention Modules (MFPA) and a Boundary-Aware Module (BAM). Input the morphological attention features into the 3D image boundary-aware segmentation module for feature extraction to obtain boundary-aware features; 3D Image Boundary-Aware Discrimination Module: Concatenate and fuse the morphological attention features and the boundary-aware features to obtain the segmentation result, then input the segmentation result into the 3D image boundary-aware discriminator for evaluation, and then feedback and optimize the 3D image boundary-aware segmentation network according to the evaluation result.

[0030] Example 3 In the verification experiment of the present invention, four experimental metrics are used to verify the effectiveness of the proposed method, namely HD95mm, DSC, Params, and FLOPs; HD95mm represents the boundary difference between the predicted segmentation image result and the ground truth label. The lower the value, the closer the boundary predicted by the model is to the true boundary; DSC (Dice Similarity Coefficient) is used to measure the overlap degree between the predicted segmentation result and the manual segmentation label. The higher the DSC value, the higher the overlap degree between the segmentation result of the model and the ground truth label; Params represents the number of parameters of the model, reflecting the complexity of the model. Under the same segmentation result, the model with fewer parameters is better; FLOPs represents the number of floating-point operations, used to measure the computational complexity of the model. The lower the FLOPs value, the higher the computational efficiency of the model.

[0031] Under the same experimental conditions and environment configurations, compare the proposed 3D medical image fuzzy boundary segmentation method based on multi-feature perception with other 3D image segmentation methods to verify the effectiveness of the proposed method. The other comparison models are as follows: U-Net, a classic image segmentation model, which realizes the extraction and fusion of multi-scale features through convolution and deconvolution operations. It is suitable for basic image segmentation tasks, but the receptive field of the model structure is limited, and its ability to segment images in complex scenarios is weak; Swin UNet, a 3D image segmentation model based on Swin Transformer, which realizes the fusion of local and global features of 3D images through the window attention mechanism, but the computational complexity of the model is relatively high; nnFormer, an improved hybrid image segmentation model, which uses an attention module to realize the feature pyramid structure. The performance of this algorithm is relatively excellent, but its computational complexity and the number of parameters are relatively high; UNETR, which uses Transformer as the encoder to capture the long-range dependence of 3D image features, and the decoder is based on a 3D convolutional network. This model is suitable for 3D image segmentation tasks, but the complexity of the model is relatively high.

[0032] The proposed 3D medical image fuzzy boundary segmentation method based on multi-feature perception was verified in the dataset. The experimental results are shown in Table 1. In the performance comparison with other 3D image segmentation models, the comprehensive performance of the method proposed in this invention is more excellent. Among them, the HD95mm index reached the lowest value of 10.57, indicating that the error between the predicted 3D object image boundary and the real boundary is the smallest. The DSC index of the proposed method is 83.64, which is significantly higher than that of other segmentation models, indicating that the segmentation accuracy of the model is the best. In the two indexes of Params and FLOPs, the scores of the model are 61.49 and 75.80 respectively. The scores of the two indexes show that the model has better performance with lower parameter quantity and calculation cost. The experimental results show that the method proposed in this invention has advantages in both accuracy and efficiency, verifying its effectiveness in 3D image segmentation tasks.

[0033] Table 1 Comparison results of the method of this invention and existing methods To more intuitively prove that this invention has a better 3D medical image fuzzy boundary segmentation effect, the method of this invention is compared with two existing methods, Swin Unet and U-Net, as Figure 3 shown. The following are three groups of comparison diagrams of the segmentation effects of the method of this invention and existing methods. By comparing the segmentation effects of the method of this invention, Swin-Unet and U-Net with the real segmentation image, it can be directly seen that the coincidence degree between the segmentation result area obtained by the method of this invention and the real segmentation area is higher. Thus, it can be further proved that the method in this invention is effective in 3D image segmentation tasks.

[0034] Although the specific implementation manners of the invention are described above in conjunction with the drawings, it is not a limitation on the protection scope of the invention. Based on the technical solutions of this invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of this invention.

Claims

1. A 3D medical image fuzzy boundary segmentation method based on multi-feature perception, characterized in that: The following steps are involved: S1. Preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; S2, constructing a medical 3D morphology capture attention module, inputting the low-dimensional feature sequence into the medical 3D morphology capture attention module, and obtaining morphology attention features; S3, construct a 3D image boundary perception segmentation network, which includes two multi-feature perception attention modules MFPA and a boundary perception module BAM, input the low-dimensional feature sequence into the 3D image boundary perception segmentation network for feature extraction, and obtain boundary perception features; S4. Construct a 3D image boundary-aware discriminator, concatenate and fuse the morphological attention features and boundary-aware features to obtain the segmentation result, and then input the segmentation result into the 3D image boundary-aware discriminator for evaluation, and then optimize the 3D image boundary-aware segmentation network based on the evaluation result feedback.

2. According to the method of 3D medical image fuzzy boundary segmentation based on multi-feature perception according to claim 1, it is characterized in that: S1 is as follows: Obtain the 3D medical image to be segmented from the existing data set. Any 3D medical image to be segmented is represented as , , Indicates height, Indicates width, Indicates depth, Indicates the number of channels; Construct a 3D medical image preprocessing module to segment the 3D medical image Input to the 3D medical image preprocessing module, where The size of the sliding window in the image Medium Slide, Image Divided into feature blocks, each feature block is represented by , , ,image Divided into The calculation process of a feature block is as follows: ， in, Indicates the height of the sliding window, represents the width of the sliding window, Represents the depth of the sliding window; Then, the feature blocks are flattened Flatten into a one-dimensional vector , the calculation formula is as follows: ， in, Indicates Feature Block After flattening, it corresponds to a one-dimensional vector , Represents a flattening operation; By linear transformation, the one-dimensional vector Projected into a low-dimensional feature space to obtain a fixed-length feature vector , , Represents the dimension after projection, and the calculation formula is as follows: ， in, represents the linear projection matrix, , represents the bias vector, ; at last, A fixed-length feature vector Combine into a feature sequence , , .

3. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 2 is characterized in that: S2 is as follows: Construct a medical 3D morphology capture attention module, which includes a normalization layer, a variable morphology capture attention module, and a convolution kernel size. The 3D convolution layer and convolution kernel size are The 3D convolution layer converts the low-dimensional feature sequence output by the 3D medical image preprocessing module Input to the medical 3D morphology capture attention module; Feature sequence First, it is normalized through the normalization layer to obtain the normalized features , and then input the normalized features into the variable form capture attention module, the normalized features After convolution kernel size is 3D convolutional layer and Activation function gets features , the calculation formula is as follows: ； Then, the features The convolution kernel size in the variable-shape capture attention module is 3D depthwise separable convolutional layer , the convolution kernel size is 3D dilated convolutional layer And the convolution kernel size is 3D convolutional layer , get the convolution feature , and then the convolution feature With features Splice and output features , and then the feature The convolution kernel size input to the variable shape capture attention module is The 3D convolution layer combines the obtained features with the normalized features Splice to obtain the output features of the variable form capture attention module , the calculation formula is as follows: ， ， ， in, Represents a splicing operation; Next, the output features of the variable form capture attention module With low-dimensional feature sequences The features are spliced ​​to obtain enhanced features after information supplementation , strengthen the characteristics After convolution kernel size 3D convolutional layer and the convolution kernel size is 3D convolutional layer , get the features ,feature With enhanced features Perform feature splicing to finally obtain the output morphological attention features of the medical 3D morphological capture attention module , the calculation formula is as follows: ， ， 。 4. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 3 is characterized in that: S3 is as follows: The two multi-feature perception attention modules MFPA are respectively the first multi-feature perception attention module and the second multi-feature perception attention module. Both multi-feature perception attention modules include a spatial attention feature capture module and a channel attention capture module. The input features are respectively passed through the spatial attention feature capture module and the channel attention capture module to obtain spatial attention features and channel attention features. The two features are then fused to obtain the output of the feature perception attention module. The boundary perception module concatenates and fuses the outputs of the two feature perception attention modules, constructs a boundary perception key feature map based on the intersection of the boundary area and the 3D true segmentation boundary image, and then calculates the boundary perception features from the concatenated features and the boundary perception key feature map.

5. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 4 is characterized in that: The specific operation of the multi-feature perception attention module is as follows: The feature sequence Input into the first multi-feature perception attention module to obtain the first multi-feature perception feature , and then the first multi-feature perception feature Input into the second multi-feature perception attention module to obtain the second multi-feature perception feature , the first multi-feature-aware attention module and the second multi-feature-aware attention module share weights of the query and the key; The specific operation process in the first multi-feature perception attention module is as follows: (1) Morphological attention features After the spatial attention feature capture module: Compute shared queries , Share key And the spatial value , the calculation formula is as follows: ， ， ， in, , , The projection matrices representing query, key, and value respectively; Then, by sharing the query and Share Key The dot product calculation is used to calculate the similarity between features and obtain the spatial attention feature map , and then the spatial attention feature map With space value Multiply them together to get the spatial attention feature , the calculation formula is as follows: ， ， in, represents the activation function, Indicates a shared key The transpose of Dimensions representing shared queries and keys; (2) Feature sequence After the channel attention feature capture module: ， ， ， in, and represents the query and key of the channel attention module, Indicates the channel value; By sharing the query and Share Key The dot product calculation is used to calculate the similarity between features and obtain the channel attention feature map , and then the channel attention feature map With channel value Multiply to get the channel attention feature , the calculation formula is as follows: ， ； (3) Spatial attention features and channel attention features Add and fuse to get the fused features , the fused features Input into the convolution kernel size in sequence The convolutional layer and the convolution kernel size is The convolutional layer The final first multi-feature perception attention module features are obtained , the calculation formula is as follows: ， ； The specific operation in the second multi-feature perception and attention module is the same as that in the first multi-feature perception and attention module. Input into the second multi-feature perception attention module to obtain the second multi-feature perception attention module feature .

6. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 5 is characterized in that: The operations in the boundary perception module are as follows: The first multi-feature aware attention module features and the second multi-feature aware attention module features Splicing and fusion to generate perceptual fusion features ; Using Canny edge detection algorithm to segment 3D medical images Corresponding real 3D medical segmentation image Extract the boundary of the target object, obtain the target boundary contour of the 3D image, and randomly select candidate boundary key points, and obtain the key point set , Indicates The iterations are randomly selected A set of boundary points, , according to the key point set Construct the corresponding boundary area , Indicates Key points, , , and Respectively represent The first randomly selected The x-axis, y-axis and z-axis coordinates of the key points; Calculate bounding area Compared with the real 3D medical segmentation image The intersection ratio , the calculation formula is as follows: ， in, Represents the optimal set of boundary points obtained after calculating the intersection ratio, including candidate boundary key points, according to the best boundary point set Construct the corresponding boundary-aware key feature map ; The key feature map Fusion of features with perception Fusion uses channel weighted operations to obtain boundary-aware features , the calculation process is as follows: ， in, represents element-wise addition, Represents channel-wise multiplication.

7. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 6 is characterized in that: The process of calculating the segmentation results is as follows: Boundary-aware features and morphological attention features Use splicing to fuse in the channel dimension to obtain fusion features , the fused features The convolution kernel size is used The convolutional layer adjusts the number of channels of the image and uses The activation function outputs the final predicted segmented 3D image , the calculation formula is as follows: ， ， in, represents the concatenation operation of features, Indicates that the convolution kernel size is The convolution operation.

8. The 3D medical image fuzzy boundary segmentation method based on multi-feature perception according to claim 7 is characterized in that: The 3D image boundary perception discriminator is as follows: If the 3D image boundary perception segmentation network finally predicts the segmented 3D image The preserved target region structure boundary and the real 3D medical segmentation image If the image segmentation result is consistent with the real image segmentation result, the 3D image boundary perception discriminator outputs a high score; otherwise, if the image segmentation result fails to match the real image segmentation result, the discriminator module outputs a low score. The discriminator loss function is designed so that the discriminator module can distinguish between high-scoring segmented images and wrong segmented images. The calculation formula is as follows: ， ， in, represents the discriminator function, represents the score output by the discriminator, Represents a splicing operation, Indicates that the convolution kernel size is The convolutional layer, represents the fully connected layer, represents the Sigmoid activation function, represents the discriminator loss function, represents the logarithmic function with base 2; The score output by the discriminator is used to optimize the 3D image boundary perception segmentation network, and the 3D image boundary loss is introduced to adjust its own parameters according to the score. The calculation formula is as follows: ， in, represents the 3D image boundary loss, Represents the base 2 logarithmic function.

9. A 3D medical image fuzzy boundary segmentation system based on multi-feature perception, executing a 3D medical image fuzzy boundary segmentation method based on multi-feature perception as claimed in any one of claims 1 to 8, characterized in that: Includes the following modules: Image preprocessing module: preprocess the 3D medical image to be segmented to obtain a low-dimensional feature sequence; Medical 3D morphology capture attention module: input the low-dimensional feature sequence into the medical 3D morphology capture attention module to obtain morphological attention features; 3D image boundary perception segmentation module: It includes two multi-feature perception attention modules MFPA and a boundary perception module BAM. The morphological attention features are input into the 3D image boundary perception segmentation module for feature extraction to obtain boundary perception features. 3D image boundary perception discriminator module: The morphological attention features and boundary perception features are spliced ​​and fused to obtain the segmentation result, and then the segmentation result is input into the 3D image boundary perception discriminator for evaluation, and then the 3D image boundary perception segmentation network is optimized based on the evaluation result feedback.

Images