Brain MRI (Magnetic Resonance Imaging) image recognition system and method fusing tensor and superpixel image features

Abstract

The invention belongs to the technical field of image recognition, and discloses a brain MRI image recognition system and method fusing tensor and superpixel image features. According to the method, a lightweight residual block fused with three-dimensional tensor attention is constructed through BTNet-TS, and a three-dimensional tensor is adopted to capture multi-dimensional lesion semantic features of brain MRI; secondly, providing a multi-scale feature grouping dense connection fusion strategy, and fusing bottom detail features and depth semantic information; and finally, designing an image convolution network based on superpixel segmentation, and deeply mining local correlation features of the focus. Therefore, efficient brain MRI image recognition is realized.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a brain MRI image recognition system and method that integrates tensor and superpixel features. Background Art

[0002] Magnetic resonance imaging (MRI), a commonly used radiological imaging technique in clinical practice, has become the preferred imaging method for diagnosing brain tumors due to its non-invasive nature, lack of ionizing radiation, and high soft tissue contrast. To improve the accuracy and efficiency of brain MRI diagnostics, deep learning-based medical image processing methods are increasingly being applied to brain MRI tumor diagnosis.

[0003] In the early days, statistical and machine learning-based methods were primarily used. In recent years, deep learning, primarily based on artificial neural networks, has continuously improved the automation and accuracy of computer-aided diagnosis. However, the scarcity of datasets and sample imbalance significantly impact diagnostic effectiveness. Sample ratios in medical image datasets are often uneven, and the acquired images contain interfering information such as noise and artifacts, which makes it difficult to fully extract key lesion features. Networks at each level typically have different spatial scales, resulting in differences in representation between features. Directly fusing lesion features from different levels can easily lead to them being overwhelmed by redundant information, affecting the accuracy of lesion detection and classification. Network structures of varying depths differ significantly in capturing image features, and the model struggles to dynamically adjust the emphasis on global and local features. Therefore, existing deep learning-based medical image classification models struggle to guarantee the accuracy and efficiency of brain MRI tumor diagnosis.

[0004] Patent publication number CN202010962134.7 discloses a system and method for recognizing severe tumor images. This system, based on convolutional neural networks and magnetic resonance imaging, uses a two-stage, multi-model approach to automatically diagnose brain tumors. By mining information, extracting features, and learning from historical diagnostic data, it enables highly accurate computer-assisted brain image data analysis, making diagnostic accuracy independent of physician experience while accelerating diagnosis. However, the convolution process struggles to effectively balance the key features and underlying details of the lesion, resulting in a slightly limited accuracy in tumor diagnosis.

[0005] Therefore, there is an urgent need for an accurate, efficient and interpretable brain MRI image recognition method to assist in the diagnosis of medical diseases. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a brain MRI image recognition system and method that integrates tensor and superpixel features to solve the problems in the prior art. To achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A brain MRI image recognition system that integrates tensor and superpixel features, including:

[0008] Image acquisition module: obtains public brain tumor medical images and divides them into proportions for model training;

[0009] Image preprocessing module: All brain MRI images acquired in the image acquisition module are cropped into images of the same size and standardized;

[0010] Improve model diagnostic efficiency module: The brain MRI processed by the image preprocessing module is used as a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is reduced through depthwise separable convolution;

[0011] The multi-layer semantic information module of brain MRI is integrated, and semantic features at different levels are sequentially and proportionally grouped and spliced ​​through dense connections. Multi-scale dilated convolution is then used to fuse the underlying detail features and deep semantic information.

[0012] Mining brain MRI local correlation feature module: divides the approximate area of ​​the brain MRI into a superpixel map, and deeply mines the local correlation features of the lesion based on the graph convolutional neural network;

[0013] Model training and lesion segmentation module: Build an experimental platform to train BTNet-TS and use the trained model to test the brain MRI to be diagnosed.

[0014] A brain MRI image recognition method that integrates tensor and superpixel map features is applied to the aforementioned brain MRI image recognition system that integrates tensor and superpixel map features, comprising the following steps:

[0015] S0: Obtain public brain tumor medical images and divide them into proportions for model training;

[0016] S1: All brain MRI images acquired in the image acquisition module are cropped into images of the same size and standardized;

[0017] S2: The processed brain MRI is converted into a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is reduced through depthwise separable convolution;

[0018] S3: Semantic features at different levels are grouped and concatenated in order and proportionally through dense connections, and multi-scale dilated convolution is used to fuse underlying detail features and deep semantic information;

[0019] S4: Divide the approximate area of ​​brain MRI into a superpixel map, and deeply mine the local correlation features of the lesion based on the graph convolutional neural network;

[0020] S5: Build an experimental platform to train BTNet-TS and use the trained model to test the brain MRI to be diagnosed.

[0021] A method for improving model efficiency, applied to the module for improving model diagnosis efficiency, comprises the following steps:

[0022] S20: Construct a three-dimensional tensor to extract multi-dimensional semantic features: define 3 groups for each residual block Tensor graph of size And according to the size of the residual block input feature The bilinear interpolation method is used to adjust its size, and the calculation formula is formula (1):

[0023]

[0024] S21: Fusing tensor features to enhance key semantic information: The original convolution operation is replaced by depth-wise separable convolution. The input feature map Input is then calculated according to the network structure of the residual block. After multiplying the three-dimensional tensor map with the input feature Input, it is added to the output of the stacked layer according to the skip connection method within the residual block. The calculation formula uses Equation (2), Equation (3) and Equation (4), where RB represents batch normalization and the activation function Relu:

[0025]

[0026] For deep networks, a three-layer bottleneck block is used to replace the original residual block, introducing The convolutional layer reduces the dimension of the middle layer, and the calculation formula is (5), (6) and (7):

[0027]

[0028] S23: Set the residual tensor selection mechanism: the input feature map X and the output feature map Out have the same shape of [B, C, M, N], and the three-dimensional feature tensor P xy The shape is [1, C, 7, 7], P xz and P yz The shape is [1,1,7,7]. After selecting the residual structure required by the model according to the input block, the input and output feature channels are used to determine the type of tensor used. After calculation by the basic residual block and the enhanced residual block respectively, the feature map Out of the residual block is output according to element-by-element superposition.

[0029] A multi-layer semantic information image fusion method is applied to the fusion brain MRI multi-layer semantic information module, comprising the following steps:

[0030] S31: Multi-level feature grouping: The i LR-3DA output feature maps to be fused are grouped into LF i The pooling layer is used to resize the dataset to a uniform size. The number of feature channels remains unchanged; then the feature maps are evenly divided according to the number of groups to be fused, and each feature map will be divided into i / C i Feature map F of the block i ; Finally, F i Different blocks in the sequence are spliced ​​together to form i groups of preliminary fused feature blocks Fa i , the calculation formula is formula (8) and formula (9:

[0031]

[0032] S32: Multi-scale feature extraction: Feature map Fa that initially integrates multi-level semantic information i Through multi-scale hole convolution calculation, feature fusion is performed again to obtain Fb i , the calculation formula is formula (10):

[0033]

[0034] Among them, pad represents the spacing parameter;

[0035] S33: Group feature fusion: The feature map Fb after preliminary fusion i After sequential concatenation, the convolution operation is repeated again to obtain the output feature map with the same size as the input feature map of the i-th group. The calculation formula is (11):

[0036]

[0037] Among them, the input feature map LF is a set of feature maps of different sizes, and the output feature map Out is in the shape of [B, C, M, N]. Multi-level feature grouping is performed according to the input features, and then feature blocks are extracted from each group and spliced ​​in sequence. After extracting the fusion features of each group using multi-scale dilated convolution, the preliminary fusion features of all groups are integrated to form the final fusion feature map Out.

[0038] A method for mining local correlation features of brain MRI, applied to the above-mentioned mining local correlation features module of brain MRI, comprises the following steps:

[0039] S41: Constructing superpixel map based on brain MRI: According to the input i-th layer brain MRI feature map, Convolution merges channels and performs superpixel segmentation to obtain a superpixel map SF containing multiple segmented areas i , then the segmented areas of the superpixel graph are used as nodes of the graph, and the adjacency relationship between the areas is used as the node association, and the adjacency matrix A is usedi Expressed; the calculation formula uses formula (12) and formula (13), where a jk Represents the connection relationship between adjacent nodes, if they are connected, it is 1, otherwise it is 0:

[0040]

[0041] S42: Select node features and standardize them: Select each region a in the segmentation result j The center coordinate value, average gray value mean and total number of pixels count are used as node features; the feature f of each dimension is calculated by formula (14) j Perform normalization to get z j ; where μ represents the mean of each feature dimension, and σ represents the variance:

[0042]

[0043] S43: Design graph convolution to extract and fuse superpixel features: Use the superpixel graph and its node features as the input of the multi-layer GCN to calculate and obtain the graph label F corresponding to the number of channels output by the LR-3DA layer. i , and compare it with the output feature map LF of the corresponding LR-3DA layer according to the element operation i Multiply to preserve the global feature GF i Passed to the next GCF-2S and guides the final tumor category diagnosis; the calculation formula uses formula (15) and formula (16):

[0044] F i =GCN(a j ), a j ∈SF i (15)

[0045] GF i =F i ·LF i (16).

[0046] The present invention has the following beneficial effects: (1) high accuracy in brain MRI image recognition; (2) high efficiency and stable performance in brain MRI image recognition; compared with the best models such as lightweight MobileNet and the latest RDNet, both Acc and F1 are improved. BTNet-TS can better balance stability and efficiency, proving its effectiveness in brain MRI image recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flow chart of the method of the present invention

[0048] Figure 2 This is a diagram of the architecture of the present invention;

[0049] Figure 3 Flowchart of lightweight residual block for integrating 3D tensor attention;

[0050] Figure 4 Dense connection fusion strategy diagram for multi-scale feature grouping;

[0051] Figure 5 Extract graph convolutional features for superpixel segmentation;

[0052] Figure 6 Confusion matrix of tumor diagnosis results of various models on the Brain Tumor MRI Dataset;

[0053] Figure 7 Confusion matrix of tumor diagnosis results of various models on the Brain Tumor Dataset;

[0054] Figure 8 Confusion matrix of tumor diagnosis results of various models on the Brain Tumor MRI Images 44Classes dataset;

[0055] Figure 9 This is a curve chart of loss changes during model training. DETAILED DESCRIPTION

[0056] The following is a combination of the embodiments of the present invention Figures 1-9 , the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0057] The present invention discloses a brain MRI image recognition system that integrates tensor and superpixel features, comprising:

[0058] Image acquisition module: obtains public brain tumor medical images and divides them into proportions for model training;

[0059] Image preprocessing module: All brain MRI images acquired in the image acquisition module are cropped into images of the same size and standardized;

[0060] Improve model diagnostic efficiency module: The brain MRI processed by the image preprocessing module is used as a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is reduced through depthwise separable convolution;

[0061] The multi-layer semantic information module of brain MRI is integrated, and semantic features at different levels are sequentially and proportionally grouped and spliced ​​through dense connections. Multi-scale dilated convolution is then used to fuse the underlying detail features and deep semantic information.

[0062] Mining brain MRI local correlation feature module: divides the approximate area of ​​the brain MRI into a superpixel map, and deeply mines the local correlation features of the lesion based on the graph convolutional neural network;

[0063] Model training and lesion segmentation module: Build an experimental platform to train BTNet-TS and use the trained model to test the brain MRI to be diagnosed.

[0064] Each module of the present invention can be implemented by computer software, and the relevant algorithms are compiled using Python language to realize the corresponding functions of each module.

[0065] The present invention also relates to a brain MRI image recognition method that integrates tensor and superpixel features, comprising the following steps:

[0066] Step S0: Image acquisition: Obtain public brain tumor medical images and divide them into proportions for model training.

[0067] Step S1: Image preprocessing: All brain MRI images acquired in step S0 are cropped into images of the same size and standardized.

[0068] Step S2: Improve the diagnostic efficiency of the model: The brain MRI preprocessed in step S1 is used as a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is effectively reduced through deep separable convolution.

[0069] Step S3: Fusion of multi-layer semantic information of brain MRI: Sequentially and proportionally group and concatenate the semantic features of different levels in S2 through dense connections, and use multi-scale dilated convolution to fuse the underlying detail features and deep semantic information.

[0070] Step S4: Mining local correlation features of brain MRI: Divide the approximate brain MRI area in S2 into a superpixel map, and deeply mine the local correlation features of the lesion based on the graph convolutional neural network.

[0071] Step S5: Model training and lesion segmentation: Build an experimental platform to train BTNet-TS, and use the trained model to test the brain MRI to be diagnosed.

[0072] It consists of lightweight residual blocks fused with 3D tensor attention (LR-3DA), graph convolutional feature extraction based on superpixel segmentation (GCF-2S), and multi-scale feature fusion strategy by GroupDenseNet (MFS-GD). Figure 2 shown.

[0073] The step S2 is a method for improving the efficiency of the model, such as Figure 3 As shown in Figure 2, the process of improving model diagnosis efficiency is as follows:

[0074] S20: Construct a three-dimensional tensor to extract multi-dimensional semantic features. To efficiently extract the key features of the input feature map, three groups are defined for each residual block. Tensor graph of size And according to the size of the residual block input feature The bilinear interpolation method is used to adjust its size. The calculation process is shown in formula (1).

[0075]

[0076] S21: Fusing tensor features to enhance key semantic information. To improve the computational efficiency of the residual map, the original convolution operation is replaced with depthwise separable convolution (DSC), and the input feature map Input is then calculated according to the network structure of the residual block. At the same time, the three-dimensional tensor map is multiplied with the input feature Input respectively, and then added to the output of the stacked layer according to the skip connection method within the residual block. In particular, the structure of the basic LR-3DA is as follows Figure 3 As shown, it mainly adopts The convolution is calculated as shown in Equations (2), (3) and (4), where RB stands for batch normalization and the activation function ReLU.

[0077]

[0078]

[0079] For the deep network, a three-layer bottleneck block is used to replace the original residual block. The structure of the enhanced LR-3DA is as follows: Figure 3 As shown. By introducing The convolutional layer reduces the dimension of the intermediate layer, which not only reduces the number of parameters but also helps the network learn more efficient feature representation, effectively alleviating the gradient vanishing and gradient exploding problems. Its calculation method is shown in Equations (5), (6) and (7).

[0080]

[0081] S23: Set the residual tensor selection mechanism. The input feature map X and the output feature map Out have the same shape of [B, C, M, N]. The three-dimensional feature tensor P xy The shape is [1, C, 7, 7], P xz and P yz The shape is [1, 1, 7, 7]. After selecting the residual structure required by the model based on the input block, the input and output feature channels are used to determine the type of tensor to use. After calculations by the base residual block (BR) and the enhanced residual block (SR), the feature map Out of the residual block is output by element-by-element superposition.

[0082] Said step S3, a multi-layer semantic information image fusion method, such as Figure 4 As shown in Figure 2, in order to better balance the semantic features of brain MRI at different levels, the process of fusing multi-layer semantic information of brain MRI is as follows:

[0083] S31: Multi-level feature grouping. First, the i LR-3DA output feature maps to be fused are grouped into LF. i The pooling layer is used to resize the dataset to a uniform size. The number of feature channels remains unchanged. Afterwards, the feature maps are evenly divided according to the number of groups to be fused, and each feature map will be divided into i / C i Feature map F of the block i Finally, F i Different blocks in the sequence are spliced ​​together to form i groups of preliminary fused feature blocks Fa i The process is shown in equations (8) and (9).

[0084]

[0085]

[0086] S32: Multi-scale feature extraction. In order to capture the information of different scales in the fused feature map and improve the robustness and expression ability of the model, the feature map Fa that initially fuses the multi-level semantic information is extracted. i Through multi-scale dilated convolution (DConv) calculation, feature fusion is performed again to obtain Fb i , the specific process is shown in formula (10).

[0087]

[0088] The pad parameter determines the distance between sampling points in the convolution kernel. Adjusting this parameter can expand the receptive field of the convolution kernel and capture a wider range of spatial information.

[0089] S33: Group feature fusion. Although multiple feature map splicing can provide richer information, it is still necessary to further explore high-level features to improve the expression and generalization capabilities of the network. Therefore, the feature map Fb i After sequential concatenation, the convolution operation is repeated again to obtain an output feature map with the same size as the input feature map of the i-th group. The process is shown in formula (11).

[0090]

[0091] The input feature maps LF consist of multiple groups of feature maps of varying sizes, and the output feature map Out has a shape of [B, C, M, N]. Multi-level feature grouping is performed based on the input features. Feature blocks are then extracted from each group and concatenated sequentially. After extracting fused features for each group using multi-scale dilated convolution, the initial fused features of all groups are integrated to form the final fused feature map Out.

[0092] The step S4 is a method for mining local correlation features of brain MRI, such as Figure 5 As shown in Figure 2, the process of mining local correlation features in brain MRI at different levels is as follows:

[0093] S41: Construct super-pixel map based on brain MRI. According to the input i-th layer brain MRI feature map, Convolution merges channels and performs superpixel segmentation to obtain a superpixel map SF containing multiple segmented areas i (a1,a2,a3,…,a j ). Afterwards, the segmented regions of the superpixel graph are used as nodes of the graph, and the adjacency relationship between regions is used as node association, and the adjacency matrix A is used. i The calculation process is shown in formulas (12) and (13), where a jk Represents the connection relationship between adjacent nodes. If they are connected, it is 1, otherwise it is 0.

[0094]

[0095]

[0096] S42: Select node features and standardize them. Although the node information in the graph is obtained through superpixel segmentation, it is still necessary to select the corresponding node features for GCN calculation. Therefore, this paper selects each region a in the segmentation result jThe center coordinate value (x, y), average gray value mean and total number of pixels count are used as node features. Due to the heterogeneity between node features, as shown in formula (14), the feature f of each dimension is j Perform normalization to get z j . Where μ represents the mean of each feature dimension, and σ represents the variance.

[0097]

[0098] S43: Design graph convolution to extract and fuse superpixel features. The superpixel graph and its node features are used as the input of the multi-layer GCN to calculate and obtain the graph label F corresponding to the number of channels output by the LR-3DA layer. i . And according to the element operation, it is compared with the output feature map LF of the corresponding LR-3DA layer i Multiply to preserve the global feature GF i It is passed to the next GCF-2S and guides the final tumor classification diagnosis. The process is shown in Equations (15) and (16).

[0099] F i =GCN(a j ), a j ∈SF i (15)

[0100] GF i =F i ·LF i (16)

[0101] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various deformations, modifications, and substitutions made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A brain MRI image recognition system that integrates tensor and superpixel features, characterized by: include: Image acquisition module: obtains public brain tumor medical images and divides them into proportions for model training; Image preprocessing module: All brain MRI images acquired in the image acquisition module are cropped into images of the same size and standardized; Improve model diagnostic efficiency module: The brain MRI processed by the image preprocessing module is used as a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is reduced through depthwise separable convolution; The multi-layer semantic information module of brain MRI is integrated, and semantic features at different levels are sequentially and proportionally grouped and spliced ​​through dense connections. Multi-scale dilated convolution is then used to fuse the underlying detail features and deep semantic information. Mining brain MRI local correlation feature module: divides the approximate area of ​​the brain MRI into a superpixel map, and deeply mines the local correlation features of the lesion based on the graph convolutional neural network; Model training and lesion segmentation module: Build an experimental platform to train BTNet-TS and use the trained model to test the brain MRI to be diagnosed.

2. A brain MRI image recognition method that integrates tensor and superpixel map features, applied to the brain MRI image recognition system that integrates tensor and superpixel map features as described in claim 1, comprising the following steps: S0: Obtain public brain tumor medical images and divide them into proportions for model training; S1: All brain MRI images acquired in the image acquisition module are cropped into images of the same size and standardized; S2: The processed brain MRI is converted into a three-dimensional tensor to capture the multi-dimensional semantic features of the lesion, and the number of model parameters is reduced through depthwise separable convolution; S3: Semantic features at different levels are grouped and concatenated in order and proportionally through dense connections, and multi-scale dilated convolution is used to fuse underlying detail features and deep semantic information; S4: Divide the approximate area of ​​brain MRI into a superpixel map, and deeply mine the local correlation features of the lesion based on the graph convolutional neural network; S5: Build an experimental platform to train BTNet-TS and use the trained model to test the brain MRI to be diagnosed.

3. A method for improving model efficiency, applied to the module for improving model diagnosis efficiency according to claim 1, characterized in that: The following steps are involved: S20: Construct a three-dimensional tensor to extract multi-dimensional semantic features: define 3 groups for each residual block Tensor graph of size And according to the size of the residual block input feature The bilinear interpolation method is used to adjust its size, and the calculation formula is formula (1): S21: Fusing tensor features to enhance key semantic information: The original convolution operation is replaced by depth-wise separable convolution. The input feature map Input is then calculated according to the network structure of the residual block. After multiplying the three-dimensional tensor map with the input feature Input, it is added to the output of the stacked layer according to the skip connection method within the residual block. The calculation formula uses Equation (2), Equation (3) and Equation (4), where RB represents batch normalization and the activation function Relu: For deep networks, a three-layer bottleneck block is used to replace the original residual block, introducing The convolutional layer reduces the dimension of the middle layer, and the calculation formula is (5), (6) and (7): S23: Set the residual tensor selection mechanism: the input feature map X and the output feature map Out have the same shape of [B, C, M, N], and the three-dimensional feature tensor P xy The shape is [1, C, 7, 7], P xz and P yz The shape is [1,1,7,7]. After selecting the residual structure required by the model according to the input block, the input and output feature channels are used to determine the type of tensor used. After calculation by the basic residual block and the enhanced residual block respectively, the feature map Out of the residual block is output according to element-by-element superposition.

4. A multi-layer semantic information image fusion method, applied to the brain MRI multi-layer semantic information fusion module according to claim 1, characterized in that: The following steps are involved: S31: Multi-level feature grouping: group the i LR-3DA output feature maps to be fused i The pooling layer is used to resize the dataset to a uniform size. And the number of feature channels remains unchanged; Then the feature maps are divided evenly according to the number of groups to be fused. Each feature map will be divided into i / C i Feature map F of the block i ; Finally, F i Different blocks in the sequence are spliced ​​together to form i groups of preliminary fused feature blocks Fa i , the calculation formula is formula (8) and formula (9: S32: Multi-scale feature extraction: Feature map Fa that initially integrates multi-level semantic information i Through multi-scale hole convolution calculation, feature fusion is performed again to obtain Fb i , the calculation formula is formula (10): Among them, pad represents the spacing parameter; S33: Group feature fusion: The feature map Fb after preliminary fusion i After sequential concatenation, the convolution operation is repeated again to obtain the output feature map with the same size as the input feature map of the i-th group. The calculation formula is formula (11): Among them, the input feature map LF is a set of feature maps of different sizes, and the output feature map Out is in the shape of [B, C, M, N]. Multi-level feature grouping is performed according to the input features, and then feature blocks are extracted from each group and spliced ​​in sequence. After extracting the fusion features of each group using multi-scale dilated convolution, the preliminary fusion features of all groups are integrated to form the final fusion feature map Out.

5. A method for mining local correlation features of brain MRI, applied to the mining local correlation feature module of brain MRI according to claim 1, characterized in that: The following steps are involved: S41: Constructing superpixel map based on brain MRI: According to the input i-th layer brain MRI feature map, Convolution merges channels and performs superpixel segmentation to obtain a superpixel map SF containing multiple segmented areas i , then the segmented areas of the superpixel graph are used as nodes of the graph, and the adjacency relationship between the areas is used as the node association, and the adjacency matrix A is used i express; The calculation formula adopts formula (12) and formula (13), where a jk Represents the connection relationship between adjacent nodes, if they are connected, it is 1, otherwise it is 0: S42: Select node features and standardize them: Select each region a in the segmentation result j The center coordinate value, average gray value mean and total number of pixels count are used as node features; the feature f of each dimension is calculated by formula (14) j Perform normalization to get z j ; where μ represents the mean of each feature dimension, and σ represents the variance: S43: Design graph convolution to extract and fuse superpixel features: Use the superpixel graph and its node features as the input of the multi-layer GCN to calculate and obtain the graph label F corresponding to the number of channels output by the LR-3DA layer. i , and compare it with the output feature map LF of the corresponding LR-3DA layer according to the element operation i Multiply to preserve the global feature GF i Passed to the next GCF-2S and guides the final tumor category diagnosis; the calculation formula uses formula (15) and formula (16): F i =GCN(a j ),a j ∈SF i (15) GF i =F i ·LF i (16)。

Images