Medical image segmentation method, system and device based on fuzzy logic modeling and dynamic graph convolutional network, and medium

Abstract

The invention relates to the field of medical image processing, in particular to a medical image segmentation method, system and device based on fuzzy logic modeling and a dynamic graph convolutional network, and a medium, and the method comprises the following steps: S1, constructing a BT-Net 3D U-shaped deep network integrated with a boundary integrity enhancement module and a trajectory continuity enhancement module; s2, extracting input features; s3, a boundary integrity enhancement module obtains enhanced boundary features based on uncertain estimation and a fuzzy neural network; the track continuity enhancement module obtains enhanced continuity imaging features based on the dynamic graph convolutional network; and S4, performing feature reconstruction to obtain a final segmentation result. According to the method, uncertainty is reduced based on fuzzy logic modeling, the integrity of the blood vessel boundary can be iteratively enhanced, so that the ability of the network to determine the blood vessel boundary under complex conditions is improved, the continuity of the blood vessel trajectory is iteratively enhanced based on the dynamic graph convolutional network, and the ability of the network to process cross-scale complex micro blood vessel structures is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a medical image segmentation method, system, device and medium based on fuzzy logic modeling and dynamic graph convolutional network. Background Art

[0002] Coronary heart disease is a common cardiovascular disease. Coronary computed tomography angiography (CCTA) is a practical non-invasive vascular imaging technology that provides an important imaging basis for the diagnosis and analysis of coronary heart disease. Coronary artery segmentation is a key step in the evaluation of coronary heart disease. In clinical practice, doctors mainly rely on manual interactive coronary artery segmentation for quantitative diagnosis, but this process is labor-intensive and time-consuming, and the results may also be subjective. Therefore, it is necessary to automatically and accurately segment the coronary artery for diagnosis.

[0003] Existing methods are often not accurate enough to meet clinical application standards because the coronary arteries have complex anatomical structures and morphologies, and exhibit unique characteristics such as sparseness and low contrast with adjacent tissues, as follows: 1. Blood vessels are long and thin structures with lumen diameters ranging from 1 mm to 5 mm, resulting in many difficult-to-segment areas and discontinuous segmentation results; 2. The image volume ratio is small (less than 0.05%), which leads to a serious imbalance between vascular and non-vascular pixels; 3. The contrast of some blood vessel locations is extremely low, making the blood vessel boundaries blurred.

[0004] In recent years, a large number of literatures have proposed methods for vessel segmentation, including classical machine learning and modern deep learning methods. A comprehensive review of coronary artery segmentation can be found in the review paper (Gharleghi et al., 2022). Early methods were mainly traditional rule-based methods, such as tracking algorithms (Moccia et al., 2018), active contour models (Lesage et al., 2016), and level set algorithms (Frangi et al., 1998), which showed promising results on limited datasets. However, these methods rely heavily on hand-designed engineering features and require rich domain knowledge and expert input, which makes it difficult to achieve ideal robustness in comprehensive coronary artery segmentation.

[0005] Unlike traditional methods that use hand-crafted features, modern deep learning methods are data-driven and can automatically extract complex image features. They have become the preferred method for medical image segmentation. U-net, especially its encoder-decoder architecture, has been widely used in medical image segmentation. Among them, CE-Net (Gu et al., 2019) uses multiple convolution branches with different receptive fields to enhance the model's ability to capture multi-scale information. CPFNet (Feng et al., 2020) uses a scale-aware layer to adaptively adjust the size of the effective receptive field; CA-Net (Gu et al., 2020) learns to make full use of multiple attentions to improve medical image segmentation performance. Zhang et al. (2021) introduced a unique 3D multi-attention-guided multi-task learning network for simultaneous gastric tumor segmentation and lymph node classification. However, these methods were not explicitly developed for coronary artery segmentation and did not fully utilize the continuity characteristics of coronary arteries. TransBTS (Wang et al., 2021) uses transformers for MRI brain tumor segmentation, uses 3D CNN to extract volumetric spatial feature maps, and uses transformers for global feature modeling. In addition, MedT (Valanarasu et al., 2021) with a gated axial attention model enhances existing designs by introducing a local-global training strategy, but vision transformers usually require a large amount of data to learn inductive biases. These methods still face challenges with low continuity and completeness in the coronary artery segmentation results, which are usually insufficient to meet clinical standards. Summary of the invention

[0006] In view of the problem in the prior art that coronary artery segmentation cannot be accurately segmented due to low continuity and integrity, the present invention provides a medical image segmentation method, system, device and medium based on fuzzy logic modeling and dynamic graph convolutional network.

[0007] The present invention is achieved through the following technical solutions: A medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network includes the following steps: S1, a BT-Net 3D U-shaped deep network with integrated boundary integrity enhancement module and trajectory continuity enhancement module based on the classic U-shaped encoder-decoder architecture; S2, taking the medical image of coronary arteries as the research object, the encoder extracts input features from the medical image; S3, the boundary integrity enhancement module is based on the encoder i The low-level features of the decoder stage iThe high-level features of the +1 stage are used as input to identify the detection area in the medical image based on uncertainty estimation, and then a trainable uncertainty reduction process is performed on the boundary area of ​​the detection area based on a fuzzy neural network to obtain enhanced boundary features; The trajectory continuity enhancement module performs graph inference on the feature graph obtained in the last stage of the encoder to construct a dynamic graph structure that reflects the association between different channels in the coronary artery structure learning process; then, it performs graph convolution network inference through a two-layer graph convolution network to obtain enhanced continuous imaging features; S4, the decoder reconstructs the continuous imaging features and the enhanced boundary features, and combines the application of convolution kernel layer and sigmoid layer on the output of the encoder to obtain the final segmentation map.

[0008] Preferably, in S1, the encoder and the decoder each include four blocks with a residual network as the main body, and each block has a residual network as the main body; The BT-Net 3DU deep network is trained by jointly optimizing the weighted cross entropy (WCE) loss and the Dice similarity coefficient (DSC) loss.

[0009] Preferably, the process of identifying the detection area in the medical image based on uncertainty estimation is: Take the low-level features of the encoder at stage i And the decoder i +1 stage advanced features As input, C , H , W and D Respectively represent the channel, height, width and depth of the input features; Advanced Features Upsampling to low-level features Same size ; After upsampling, the convolutional layer and sigmoid activation function ( σ ) Generate a predicted probability map from the upsampled features : ; Based on the predicted probability map Computing uncertainty maps :

[0010] in, is the predicted probability map The value of the jth position in , is the threshold value; Then, the uncertainty map Multiply and add the result back to In the process of generating enhanced features :

[0011] in, represents element-wise multiplication, Represents element-wise addition.

[0012] Preferably, the specific steps of performing a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network to obtain enhanced boundary features are: First, to enhance the features As input, the LeakyReLU activation function is used to process the data and generate activation features , C, H, W and D represent the channel, height, width and depth of the input feature respectively; then each feature point in each channel c is processed by M Gaussian membership functions to convert the feature map into a fuzzy representation, and then the fuzziness of each channel c is calculated using the "OR" logic and the fuzziness tensor , and finally the fuzzy logic information With enhanced features The residual operation is used to integrate and fuse the integrated features with the upsampled high-level features to generate the refined feature map of the boundary integrity enhancement module. ;in, They represent the mean and standard deviation of the Gaussian membership function, respectively, and are randomly initialized learnable parameters; M is a constant with a value of 4.

[0013] Preferably, in S3, the specific steps of graph derivation are as follows: First, the feature graph extracted by the last stage E5 of the encoder is (Reshaped into the node feature matrix in the graph , where C, H, W and D represent the number of feature map channels, height, width and depth respectively C=256, represents the number of nodes V; each node corresponds to a channel in feature X and has a feature vector ; At this time, the graph data contains 256 nodes, and the feature vector dimension of each node is ; At the same time, in order to obtain the adjacency matrix representing the edge weights between graph nodes , to model the relationship between channels. The specific steps are as follows: First, the feature map F is converted into a 256-dimensional column vector through the channel attention mechanism , and its calculation formula is as follows:

[0014] in, is the Sigmoid function, FC represents the fully connected layer, ReLU represents the rectified linear unit, and GAP represents the global average pooling; The second step is to transform the column vector Input to the adjacency matrix activation function To generate the adjacency matrix A:

[0015] in, and They are The i-th and j-th elements of Is a continuous activation function with a value range between 0 and 1, which is used to generate the channel weight vector Mapped to a matrix with diagonal elements equal to 1, the matrix elements represent the edge weights between channel nodes ; The third step is to introduce a learnable parameter matrix To dynamically adjust the connection weights of the adjacency matrix:

[0016] The dynamic adjacency matrix It is continuously updated during training, reflecting the changes in the relationships between channels as the network learns the vascular structure; So far, the constructed graph data is represented as G = (V, E, X, A), where V represents the nodes of the graph, E represents the edges of the graph, X is the node feature matrix, and A is the dynamic adjacency matrix.

[0017] Preferably, the specific steps of graph convolutional network reasoning are as follows: First, define a two-layer GCN network and perform graph convolution operations on the obtained graph data G:

[0018] in, is the output of the lth layer, LeakyReLU is the activation function, W is the learnable weight matrix, and X is the feature matrix of the nodes in the graph data G. To alleviate the problem of gradient disappearance or explosion, the original adjacency matrix A is replaced by , and its calculation formula is:

[0019] in, is the identity matrix yes The degree matrix of .

[0020] After the graph convolution operation, the enhanced feature representation will be mapped back to the original latent space, represented as ; Finally, the feature map It is fused with the high-resolution feature map from the encoder E4 layer and passed to the decoding layer D4 of the network.

[0021] Preferably, in S4, the kernel size of the convolution kernel layer is 1×1×1.

[0022] A medical image segmentation system based on fuzzy logic modeling and dynamic graph convolutional network includes a decoder, an encoder, a boundary integrity enhancement module and a trajectory continuity enhancement module, wherein the encoder is used to extract image features; the boundary integrity enhancement module includes an uncertainty estimation module and a fuzzy attention derivation module, wherein the uncertainty estimation module is used to process the input features of each stage in the BT-Net 3D U-shaped deep network architecture, and identify the detection area in the medical image based on uncertainty estimation; the fuzzy attention derivation module is used to further refine the enhanced features obtained from the uncertainty estimation module, and perform a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network to obtain enhanced boundary integrity features; the trajectory continuity enhancement module includes a graph derivation module and a graph convolutional network reasoning module, wherein the graph derivation module is used to process the feature map obtained in the last stage of the encoder in the BT-Net 3D U-shaped deep network architecture to construct a dynamic graph structure reflecting the association between different channels in the coronary artery structure learning process; the graph convolutional network reasoning module is used to dynamically reason the dynamic graph structure through a two-layer graph convolutional network to obtain enhanced continuity features.

[0023] A terminal device comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the steps of the method are implemented when the processor executes the computer program.

[0024] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0025] Compared with the prior art, the present invention has the following beneficial effects: The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network of the present invention uses BT-Net 3D network for coronary artery segmentation, reduces uncertainty based on fuzzy logic modeling through boundary integrity enhancement (BCE) module, can iteratively enhance the integrity of vascular boundaries, thereby improving the network's ability to determine vascular boundaries in complex situations, and proposes a trajectory continuity enhancement (TCE) module based on dynamic graph convolutional network (GCN), which can iteratively enhance the continuity of vascular trajectories, thereby improving the network's ability to handle complex microvascular structures across scales. A large number of experiments have shown that this method not only performs well in coronary artery segmentation tasks, but also shows good segmentation performance on other vascular types, and surpasses many advanced algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of the medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network of the present invention; Figure 2 It is a structural diagram of a BT-Net 3D U-shaped deep network in a medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network of the present invention; Figure 3 It is a structural diagram of a boundary integrity enhancement module in a medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network of the present invention.

[0027] Figure 4 It is a structural diagram of a trajectory continuity enhancement module in a medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network of the present invention; Figure 5 is the coronary artery segmentation result on the CAS-119 dataset in the embodiment; Figure 6 It is the coronary artery segmentation result on the ASOCA dataset in the embodiment. DETAILED DESCRIPTION

[0028] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.

[0029] The present invention discloses a medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network. Figure 1 , including the following steps: S1, a BT-Net 3D U-shaped deep network integrated with a boundary integrity enhancement module and a trajectory continuity enhancement module is constructed based on the classic U-shaped encoder-decoder architecture.

[0030] Specifically, refer to Figure 2, the BT-Net 3D U-shaped deep network is based on the classic U-shaped encoder-decoder architecture, which can explore the feature enhancement of vascular boundary integrity and trajectory continuity. At the same time, the BT-Net 3D U-shaped deep network is trained by jointly optimizing the weighted cross entropy (WCE) loss and the Dice similarity coefficient (DSC) loss. Among them, the WCE loss can adjust the learning bias between blood vessels and background during training, while the DSC loss can ensure the segmentation of small blood vessels. For the binary segmentation task, the definitions of WCE loss and DSC loss are as follows:

[0031]

[0032] Among them, the parameters is the class weight of the vascular structure, which can be used to estimate the probability of the class of all voxels get:

[0033] N is the number of voxels, and Represent the predicted probability and true value of the i-th voxel, respectively. is the minimum smoothing term to prevent the denominator from being 0, and is set to 0.0001 in the experiment. Inspired by the adaptive weighted loss function (CA et al., 2020), adaptive weight update is used to combine the WCE loss and the DSC loss. Therefore, the total joint optimization loss function of the BT-Net 3D U-shaped deep network proposed in this invention is:

[0034] Among them, the parameters is the balance parameter between WCE loss and DSC loss, empirically set to Auxiliary items It can be regarded as a weight learning item. Parameter and are trainable, meaning that only initial values ​​need to be provided and then updated as the network is trained. During training, they interact and constrain each other, so an automatic balance occurs.

[0035] S2, taking the medical image of the coronary artery as the research object, the encoder extracts input features from the medical image.

[0036] S3, the boundary integrity enhancement module is based on the encoder i The low-level features of the decoder stage iThe high-level features of the +1 stage are used as input to identify the detection area in the medical image based on uncertainty estimation, and then a trainable uncertainty reduction process is performed on the boundary area of ​​the detection area based on a fuzzy neural network to obtain enhanced boundary features; The trajectory continuity enhancement module performs graph inference on the feature graph obtained in the final stage of the encoder to construct a dynamic graph structure that reflects the association between different channels during the coronary artery structure learning process; subsequently, a graph convolutional network inference process is performed through a two-layer graph convolutional network to obtain enhanced continuity imaging features.

[0037] Reference Figure 3 In order to enhance the ability of the network architecture to segment vascular boundaries in complex situations, a boundary completeness enhancement module (BCE module) is introduced to replace the skip connections between the encoder and decoder from stage 1 to stage 4. The BCE module first performs boundary uncertainty estimation to identify regions that are particularly challenging for the segmentation network, and then performs a trainable uncertainty reduction process on these detected boundary regions through a custom-designed fuzzy neural network to enhance their imaging features.

[0038] The BCE module includes an uncertainty estimation module and a fuzzy attention derivation module. These components work together to iteratively refine the segmentation results and reduce uncertainty. Among them, the uncertainty estimation module is used to process the input features of each stage in the U-shaped network architecture, specifically: Encoder No. i Low-level features of the stage And the decoder i +1 stage advanced features As input, C, H, W, and D represent the channel, height, width, and depth of the input features, respectively; Advanced Features Upsampling to low-level features Same size, ; After upsampling, the convolution layer (Conv) and the sigmoid activation function ( ) Generate a predicted probability map from the upsampled features : ; To capture the uncertainty of the model, the predicted probability plot Computing uncertainty maps :

[0039] in, is the predicted probability map The value of the jth position in , is the threshold; through this formula, the uncertain areas in the predicted probability map can be identified. Studies have shown (Zhang et al., 2020; Kim et al., 2021) that boundary areas are usually associated with fuzzy probability scores close to 0.5. Therefore, setting The uncertainty map is 0.5, which is aligned with some uncertain regions. By focusing on the areas where the network confidence is closest to this threshold, the uncertainty map effectively highlights the areas that need more attention, guiding the network to focus on these challenging areas.

[0040] Then, the uncertainty map Multiply and add the result back to In the process of generating enhanced features :

[0041] in, represents element-wise multiplication, Represents element-wise addition.

[0042] The fuzzy attention derivation module is used to further refine the enhanced feature representation obtained from the uncertainty estimation module, that is, to perform a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network. The specific steps for obtaining enhanced imaging features are: First, to enhance the features As input, the LeakyReLU activation function is used to process the data and generate activation features , C, H, W and D represent the channel, height, width and depth of the input feature respectively; then each feature point in each channel c is processed by M (M is a constant, and the value is 4 in this invention) Gaussian membership functions to convert the feature map into a fuzzy representation. This process assigns a fuzzy language term label to each feature point according to the feature map and the fuzzy kernel. Gaussian membership functions are used to smoothly represent boundary information, enabling the model to capture and emphasize uncertain areas while retaining key structural details.

[0043] The membership function is defined as:

[0044] in, m = 1, 2, ..., M, c = 1, 2, ..., C, and and They are mThe mean and standard deviation of the Gaussian membership functions are randomly initialized learnable parameters. By converting the feature map into a fuzzy representation, the model is able to use non-exclusive fuzzy membership values ​​to handle noise and uncertainty in the image, which is more robust than deterministic features.

[0045] Then use the "OR" logical aggregation operation to calculate the ambiguity of each channel c and the fuzziness tensor ,The “OR” logical aggregation operation method helps to construct important fuzzy feature representations that can ,retain target features while eliminating irrelevant features.

[0046] Among them, the blur tensor of the c channel ( ) is calculated as:

[0047] in, Represent and operate to obtain the fuzziness tensor Corresponding to the original input feature map The same dimensions, Represent the mean and standard deviation of the Gaussian membership function, respectively, and are randomly initialized learnable parameters.

[0048] Finally, the fuzzy logic information With enhanced features The residual operation is used to integrate and fuse the integrated features with the upsampled high-level features to generate the refined feature map of the BCE module. :

[0049] Refined feature map It is then input to the next stage of the decoder.

[0050] As shown above, the BCE module operates at each stage of the network, implicitly implementing an iterative refinement design through its integration in the U-shaped architecture. This design gradually enhances boundary integrity and reduces uncertainty through multiple stages, thereby significantly improving the accuracy and completeness of vessel boundary segmentation and producing more precise and detailed results.

[0051] The trajectory continuity enhancement module (TCE module) is introduced into the BT-Net 3D U-shaped deep network to improve the continuity of segmented blood vessels and reduce fragmentation. The TCE module performs graph derivation on the feature map obtained in the last stage of the encoder to construct a dynamic graph structure that reflects the association between different channels in the coronary artery structure learning process; then, the graph convolution network reasoning process is performed through a two-layer graph convolution network to obtain enhanced continuity features.

[0052] Among them, Figure 4 As shown in the figure, the specific steps of the derivation are as follows: the output feature map from encoder E5 (C = 256), construct a graph data G to represent the association relationship between different channels of the feature graph, G = (V, E, X, A), where V represents the nodes of the graph, each channel corresponds to a node; E represents the edge of the graph, indicating the relationship between nodes; X is the node feature matrix; A is the adjacency matrix, encoding the relationship between channels.

[0053] Each node Corresponds to a channel in the feature map and has a feature vector The feature map F is reshaped into ,in , represents the node feature vector of the corresponding node V in the graph data. At this time, the graph data G has 256 nodes, and the dimension of the feature vector of each node is .

[0054] At the same time, obtain the adjacency matrix representing the edge weights between graph nodes in the feature graph F , to model the relationship between channels, the specific steps are as follows: The first step is to use channel attention to convert the feature map Convert to 256-dimensional column vector , which determines the importance of each channel, and the formula is as follows:

[0055] in is the sigmoid function, FC is the fully connected layer, ReLU means rectified linear unit, and GAP means global average pooling.

[0056] Step 2: Column vector is input to the adjacency matrix activation function to obtain the adjacency matrix A, which defines the edge weights between the graph nodes:

[0057] in, and They are The i-th and j-th elements of . Function Is a continuous activation function with a value range between 0 and 1, which is used to generate the channel weight vector Mapped to a matrix with diagonal elements equal to 1, the matrix elements represent the edge weights between the corresponding channel nodes .

[0058] The third step is to transform the learnable parameter matrix Multiply , to dynamically adjust the connection weights of the adjacency matrix:

[0059] The dynamic adjacency matrix ,This adjacency matrix A evolves during training, reflecting the changes in the relationships between channels as the network learns the vascular structure; So far, the constructed graph data is represented as G = (V, E, X, A), where V represents the nodes of the graph, and each channel corresponds to a node; E represents the edges of the graph, indicating the relationship between nodes; X is the node feature matrix; A is the dynamic adjacency matrix, which encodes the relationship between channels.

[0060] In order to perform convolution operations on graph data G, researchers (Kitrungrotsakul et al., 2017; Zhou et al., 2018) proposed spectral filtering on the graph. Therefore, in the present invention, the specific steps of graph convolutional network reasoning are as follows: First, a two-layer graph convolutional neural network (GCN network, similar to Li et al., 2021; Chen et al., 2020) is defined to perform a convolution operation on the obtained graph data G, as follows:

[0061] in, is the output of layer l, LeakyReLU is the activation function, is a learnable weight matrix, and X is the feature matrix of the nodes in the graph data G. To alleviate the problem of gradient disappearance or explosion, the original adjacency matrix A is replaced by Ã, and its calculation formula is:

[0062] in, is the identity matrix, yes The degree matrix of .

[0063] After the GCN operation, the enhanced feature representation is mapped back to its original latent space , whose dimension is consistent with the input feature F. In order to preserve the global topological information and local spatial details, the enhanced feature map It is fused with the high-resolution feature map from the earlier encoding layer (i.e., E4) through the BCE module and then passed to the decoding layer D4 of the network. Embedding the TCE module at the bottom of the network significantly enhances the feature representation capability.

[0064] The feature map It is fused with the high-resolution feature map from the E4 encoding layer and then passed to the subsequent stages of the network.

[0065] S4, the decoder reconstructs the continuous imaging features and the enhanced boundary features, and combines the 1×1×1 convolution kernel layer and sigmoid layer applied on the output of the encoder to obtain the final segmentation map.

[0066] This paper proposes a BT-Net 3D deep network for high-quality coronary artery segmentation. By integrating the BCE module and the TCE module, the BT-Net3D deep network synergistically enhances vascular features and improves boundary integrity and trajectory continuity. Among them, the BCE module reduces uncertainty through fuzzy logic modeling and improves the integrity of vascular boundaries, which is used to enhance the ability of the BT-Net 3D deep network to determine vascular boundaries in complex situations; the TCE module dynamically evaluates the spatial context of vascular segments based on a dynamic graph convolutional network (GCN), provides adaptive perception, improves the continuity of vascular trajectories, and enhances the network's ability to handle complex and subtle vascular structures. Extensive experiments conducted on internal and public datasets show that the proposed method performs well in segmentation performance and generalization ability, and is easy to migrate to the vascular segmentation problem of other organs, surpassing a variety of advanced algorithms.

[0067] Build a graph data G that represents the association between different channels of the feature graph, where G = (V, E, X, A), V represents the nodes of the graph, and each channel corresponds to a node; E represents the edge of the graph, which represents the relationship between nodes; X is the node feature matrix, and A is the adjacency matrix, which encodes the relationship between channels; each node Corresponds to a channel in the feature map and has a feature vector .

[0068] Experiments are conducted based on the CTA119 dataset and the ASOCA dataset to evaluate the performance of the proposed method in the coronary artery segmentation task. The CTA119 dataset collects CCTA volume data of 119 subjects, with a voxel size of 512 × 512 × N (N ∈ [199, 352]) and a resolution of ). All images were scanned by certified radiographers in the Department of Cardiovascular Medicine of a top hospital (class tertiary hospital) in a certain country and were approved by the ethics committee. The annotation of the coronary arteries was obtained in the following way: First, annotation was performed by three experienced experts. Second, the three experts collectively checked the annotation quality of each case. If adjustments were needed according to the judgment of the expert group, they were modified and improved.

[0069] ASOCA (Automated Coronary Artery Segmentation Dataset) is a public dataset from the MIC Coronary Artery I 2020 Challenge, which aims to segment the coronary arteries (Gharleghi et al., 2023). The dataset contains 40 CCTA images with simple and clear annotations of the main branches, 20 of which are healthy patients and the other 20 are patients with coronary artery disease. The image size is 512 × 512 × (168 to 334) voxels, the in-plane resolution ranges from 0.32 to 0.46 mm², and the inter-slice spacing is 0.625 mm.

[0070] Three types of metrics are used to quantitatively evaluate the performance of the proposed method. First, two voxel-level metrics are used: Dice Similarity Coefficient (DSC) and True Prediction Rate (TPR). It should be noted that in the binary segmentation task, DSC is mathematically equivalent to the F1 score (Grøvik et al., 2020). The values ​​of these metrics range from 0 to 1, and larger values ​​indicate higher segmentation accuracy. The calculation formula is as follows:

[0071]

[0072] Among them, P represents the predicted mask, G is the true label, Represents an intersection operation.

[0073] Then, a distance-based metric is used: the average symmetric surface distance (ASSD), which is measured in millimeters. This metric reflects the boundary error in the segmentation result and is calculated as follows:

[0074] Where P and G represent the predicted mask and true label, respectively, and Q(P) and Q(G) represent the surface voxel sets of P and G, respectively. The smaller the ASSD value, the higher the segmentation accuracy.

[0075] In order to evaluate the specific characteristics of tubular structures, the clDice metric (Shit et al., 2021) was introduced to evaluate the segmentation results to preserve the vascular structure and connectivity. This metric complements existing metrics such as DSC, TPR, and ASSD, which focus on topological accuracy.

[0076] Finally, the Wilcoxon signed-rank test was used to calculate the p-values ​​of DSC for different models, and p < 0.05 was considered statistically significant. Lower ASSD values ​​and higher values ​​of other indicators indicate better segmentation results. The reported results are based on the mean and standard deviation of the test set.

[0077] The datasets used in the experiments were normalized to reduce irrelevant information and increase image contrast. First, the intensity of all voxels in the volume was clipped to a specific range. For example, the range of CAS-119 and ASOCA datasets is [-250, 450] HU. Then, the original values ​​were mapped to the range of [0, 255].

[0078] The PyTorch package was used to implement it on the NVIDIA RTX 3090 GPU. The Adam optimizer (Kingma and Ba, 2014) was used, the initial learning rate was set to 0.001, and the cosine annealing learning rate scheduler was used to gradually reduce the learning rate. In addition, the batch size was set to 2, and the training was stopped when the learning rate was lower than 10-8 or reached 600 epochs.

[0079] During training, all datasets were split into training and validation sets using five-fold cross validation. Each dataset was divided into five consecutive folds, each fold was validated once, and the remaining four folds formed the training set. In addition, the training dataset was enlarged by online data augmentation (such as rotation, flipping, and Gaussian blur).

[0080] As training input, sub-volumes of size 160 × 160 × 128 are randomly cropped. In the test phase, a sliding window of the same size is used and moved with a half-window step to cover the entire image volume. Finally, the prediction results of different sub-volumes are integrated to obtain the final result.

[0081] The proposed method is compared with 15 latest methods, which are divided into four groups: The first group contains three general segmentation networks, namely UNet3D (Çiçek et al., 2016), VNet (Milletari et al., 2016), and HighResNet3D (Li et al., 2017).

[0082] The second group contains transformer-based models, namely MedT (Valanarasu et al., 2021), TransBTS (Wang et al., 2021), and UNETR (Hatamizadeh et al., 2022).

[0083] The third group contains four general medical image segmentation methods, namely DenseVoxelNet (Yu et al., 2017), CA-Net (Gu et al., 2020), MTLN (Zhang et al., 2021), and CPFNet (Feng et al., 2020).

[0084] The last group consists of five methods specifically designed for vessel segmentation, namely Livne et al. (Livne et al., 2019), Tetteh et al. (Tetteh et al., 2020), CS2Net (Mou et al., 2021), Dang et al. (Dang et al., 2022), and MsCAS (Dong et al., 2022).

[0085] Table 1 shows the quantitative comparison results of CAS-119 and ASOCA datasets on four indicators: DSC, TPR, ASSD and clDice. All evaluation indicators in Table 1 are obtained by five-fold cross validation, using an 80:20 ratio. For the CCTA dataset, the proposed method outperforms all the compared methods. Specifically, the proposed method achieves 86.27%, 92.31%, 0.79 mm, 92.37%, 95.16% and 94.03% on DSC, TPR, ASSD and clDice, respectively. Compared with the classic UNet3D, DSC, TPR, ASSD and clDice are improved by about 4.62%, 4.15%, 0.77 mm and 5.19%, respectively, indicating the effectiveness of the proposed BT-Net method.

[0086] The best performing methods in the other three categories are UNETR, CPFNet, and MsCAS. Our method still has significant advantages over these three methods in all indicators, especially in the ASSD indicator reflecting boundary integrity and the TR, TP, and clDice indicators reflecting trajectory continuity. In addition, compared with other methods, the p-value of the proposed method on the DSC indicator is less than 0.001, indicating that the proposed method significantly outperforms other methods in segmentation performance. On the clearly labeled ASOCA dataset, the proposed method also outperforms all other comparative methods. Specifically, compared with the second best method on each indicator, the proposed method is 2.11% higher than CS2Net on DSC, 1.62% higher than CPFNet on TPR, 0.21 mm lower than MsCAS on ASSD, and 2.38% higher than CS2Net on clDice, which shows that the proposed method can comprehensively enhance the boundary integrity and trajectory continuity of coronary artery features.

[0087] Table 1 Quantitative comparison of the latest methods on the CTA119 dataset and the ASOCA dataset

[0088] Reference Figure 5 , 6 , shows the coronary artery segmentation results of different methods on the collected CAS-119 dataset and the public ASOCA dataset. In most cases, the quality of the vessel mask generated by the proposed method is better than other state-of-the-art algorithms. Figure 5 and Figure 6 As shown, the comparison methods usually show under-segmentation or over-segmentation, especially in coronary arteries with complex and tortuous structures. In contrast, the overall segmentation results of the proposed method maintain the complete and fine coronary artery structure.

[0089] from Figure 5 It can be seen that compared with other comparative methods, BT-Net can capture the details and boundaries of coronary arteries more accurately. For example, classic networks such as UNet3D, VNet and HighResNet3D usually have more missed detections when segmenting small branches of coronary arteries, while the BT-Net method proposed in the present invention can maintain complete vascular branches and effectively avoid under-segmentation and over-segmentation problems. In addition, although the transformer series methods (such as UNETR and TransBTS) show certain advantages in the segmentation of large structures, they still perform poorly in the continuity and accuracy of small vascular branches. BT-Net, due to the introduction of GCN and fuzzy logic mechanisms, can better preserve the topological structure and boundary details of blood vessels.

[0090] from Figure 6 It can be seen that the ASOCA dataset is clearly labeled and can be used to evaluate the performance of the segmentation model. Compared with other methods, BT-Net can generate more detailed segmentation results while maintaining the main trunk and branches of the coronary arteries. For example, methods such as CS2Net, MsCAS, and CPFNet may experience breakage or poor continuity when dealing with coronary artery branches, while BT-Net maintains the integrity of blood vessels in all areas, especially in branching and bending areas. In addition, BT-Net can not only capture the thicker vascular trunks, but also effectively segment the thinner vascular branches, proving that it has a stronger adaptability to complex structures and small blood vessels.

[0091] In summary, the proposed method shows superior performance in qualitative results. It can not only accurately segment large structures, but also show significant advantages in processing small blood vessels and complex topological structures. This shows that BT-Net has strong robustness and generalization ability, and can stably output high-quality segmentation results under different data sets and different vascular structures. The application of this method in other complex anatomical structure segmentation tasks and the integration of more global and local contextual information can improve the performance of the network when processing large-scale 3D image data sets.

[0092] The present invention also discloses a medical image segmentation system based on fuzzy logic modeling and dynamic graph convolutional network, comprising a decoder, an encoder, a boundary integrity enhancement module and a trajectory continuity enhancement module, wherein the encoder is used to extract image features; the boundary integrity enhancement module comprises an uncertainty estimation module and a fuzzy attention derivation module, wherein the uncertainty estimation module is used to process the input features of each stage in the BT-Net 3D U-shaped deep network architecture, and identify the detection area in the medical image based on uncertainty estimation; the fuzzy attention derivation module is used to further refine the enhanced features obtained from the uncertainty estimation module, and perform a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network to obtain enhanced boundary integrity features; the trajectory continuity enhancement module comprises a graph derivation module and a graph convolutional network reasoning module, wherein the graph derivation module is used to process the feature graph obtained in the last stage of the encoder in the BT-Net 3D U-shaped deep network architecture to construct a dynamic graph structure reflecting the association between different channels in the coronary artery structure learning process; the graph convolutional network reasoning module performs dynamic reasoning on the dynamic graph structure through a two-layer graph convolutional network to obtain enhanced continuity features.

[0093] The present invention also discloses a terminal device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, and the steps of the above method are implemented when the processor executes the computer program.

[0094] The present invention also discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to impose any limitation on the technical solution of the present invention. Those skilled in the art should understand that, without departing from the spirit and principles of the present invention, the technical solution can also be subjected to several simple modifications and substitutions, and these modifications and substitutions are also within the scope of protection covered by the claims.

Claims

1. A medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network, characterized in that: The following steps are involved: S1, a BT-Net 3D U-shaped deep network with integrated boundary integrity enhancement module and trajectory continuity enhancement module based on the classic U-shaped encoder-decoder architecture; S2, taking the medical image of coronary arteries as the research object, the encoder extracts input features from the medical image; S3, the boundary integrity enhancement module is based on the encoder i The low-level features of the decoder stage i The high-level features of the +1 stage are used as input to identify the detection area in the medical image based on uncertainty estimation, and then a trainable uncertainty reduction process is performed on the boundary area of ​​the detection area based on a fuzzy neural network to obtain enhanced boundary features; The trajectory continuity enhancement module performs graph inference on the feature graph obtained in the last stage of the encoder to construct a dynamic graph structure reflecting the association between different channels in the coronary artery structure learning process, and then performs a graph convolution network inference process through a two-layer graph convolution network to obtain enhanced continuous imaging features; S4, the decoder reconstructs the continuous imaging features and the enhanced boundary features, and combines the application of convolution kernel layer and sigmoid layer on the output of the encoder to obtain the final segmentation map.

2. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 1, characterized in that: In S1, both the encoder and decoder contain four blocks with residual networks as the backbone, and each block has a residual network as the backbone; The BT-Net 3D U-shaped deep network is trained by jointly optimizing the weighted cross entropy loss and the Dice similarity coefficient loss.

3. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 1, characterized in that: In S3, the process of identifying the detection area in the medical image based on uncertainty estimation is: Encoder No. i Low-level features of the stage And the decoder i +1 stage advanced features As input, C , H , W and D Represent the channel, height, width and depth of the input features respectively; Advanced Features Upsampling to low-level features Same size ; After upsampling, the convolutional layer and sigmoid activation function ( σ ) Generate a predicted probability map from the upsampled features : ； Based on the predicted probability map Computing uncertainty maps : in, is the predicted probability map The value of the jth position in , is the threshold value; Then, the uncertainty map Multiply and add the result back to In the process of generating enhanced features : in, represents element-wise multiplication, Represents element-wise addition.

4. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 3, characterized in that: In S3, the specific steps of performing a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network to obtain enhanced imaging features are: First, to enhance the features As input, the LeakyReLU activation function is used to process the data and generate activation features , C, H, W and D represent the channel, height, width and depth of the input feature respectively; then each feature point in each channel c is processed by M Gaussian membership functions to convert the feature map into a fuzzy representation, and then the fuzziness of each channel c is calculated using the "OR" logic and the fuzziness tensor , and finally the fuzzy logic information With enhanced features The residual operation is used to integrate and fuse the integrated features with the upsampled high-level features to generate the refined feature map of the boundary integrity enhancement module. ;in, They represent the mean and standard deviation of the Gaussian membership function, respectively, and are randomly initialized learnable parameters; M is a constant.

5. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 4, characterized in that: In S3, the specific steps of graph derivation are as follows: First, the feature graph extracted by the last stage E5 of the encoder Reshape into a node feature matrix in the graph , where C=256, , represents the number of nodes V, each node corresponds to a channel in feature X and has a feature vector ; At this time, the graph data contains 256 nodes, and the feature vector dimension of each node is ; At the same time, obtain the adjacency matrix representing the edge weights between graph nodes in the feature graph F , to model the relationship between channels. The specific steps are as follows: First, the feature map F is converted into a 256-dimensional column vector through the channel attention mechanism , and its calculation formula is as follows: in, is the Sigmoid function, FC represents the fully connected layer, ReLU represents the rectified linear unit, and GAP represents the global average pooling; Step 2: Column vector is input to the adjacency matrix activation function To generate the adjacency matrix A: in, and They are The i-th and j-th elements of ; function Is a continuous activation function with a value range between 0 and 1, which is used to generate the channel weight vector Mapped to a matrix with diagonal elements equal to 1, the matrix elements represent the edge weights between channel nodes ; The third step is to introduce a learnable parameter matrix To dynamically adjust the connection weights of the adjacency matrix: The dynamic adjacency matrix ,The adjacency matrix A is continuously updated during the training process, reflecting the changes in the relationships between channels when the network learns the vascular structure; So far, the constructed graph data is represented as G = (V, E, X, A), where V represents the nodes of the graph, and each channel corresponds to a node; E represents the edges of the graph, indicating the relationship between nodes; X is the node feature matrix; A is the dynamic adjacency matrix, which encodes the relationship between channels.

6. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 4, characterized in that: In S3, the specific steps of graph convolutional network reasoning are as follows: First, define a two-layer graph convolutional neural network and perform graph convolution operations on the obtained graph data G: in, is the output of the lth layer, LeakyReLU is the activation function, W is the learnable weight matrix, and X is the feature matrix of the nodes in the graph data G; to alleviate the problem of gradient disappearance or explosion, the original adjacency matrix A is replaced by , and its calculation formula is: in, is the identity matrix, yes The degree matrix of After the graph convolution operation, the enhanced feature representation will be mapped back to the original latent space, represented as ; Finally, the feature map It is fused with the high-resolution feature map from the encoder E4 layer and passed to the decoding layer D4 of the network.

7. The medical image segmentation method based on fuzzy logic modeling and dynamic graph convolutional network according to claim 5, characterized in that: In S4, the kernel size of the convolution kernel layer is 1×1×1.

8. A medical image segmentation system based on fuzzy logic modeling and dynamic graph convolutional network, characterized in that: It includes a decoder, an encoder, a boundary integrity enhancement module and a trajectory continuity enhancement module, and the encoder is used to extract image features; The boundary integrity enhancement module includes an uncertainty estimation module and a fuzzy attention derivation module. The uncertainty estimation module is used to process the input features of each stage in the BT-Net 3D U-shaped deep network architecture and identify the detection area in the medical image based on uncertainty estimation. The fuzzy attention derivation module is used to further refine the enhanced features obtained from the uncertainty estimation module and perform a trainable uncertainty reduction process on the boundary area of ​​the detection area based on the fuzzy neural network to obtain enhanced boundary integrity features. The trajectory continuity enhancement module includes a graph inference module and a graph convolutional network reasoning module. The graph inference module is used to process the feature graph obtained in the final stage of the encoder in the BT-Net 3D U-shaped deep network architecture to construct a dynamic graph structure that reflects the association between different channels in the coronary artery structure learning process; the graph convolutional network reasoning module dynamically infers the dynamic graph structure through a two-layer graph convolutional network to obtain enhanced continuity features.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images