A brain MRI image recognition system and method fusing tensor and superpixel map features

Abstract

The application belongs to the technical field of image recognition, and discloses a brain MRI image recognition system and method fusing tensor and superpixel graph features. The application adopts a three-dimensional tensor to capture multi-dimensional lesion semantic features of brain MRI by constructing a light residual block fusing three-dimensional tensor attention through BTNet-TS. Then, a multi-scale feature grouping dense connection fusion strategy is proposed to fuse bottom layer detail features and deep semantic information. Finally, a graph convolution network based on superpixel segmentation is designed to deeply mine lesion local correlation features. In this way, efficient brain MRI image recognition is realized.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a brain MRI image recognition system and method that integrates tensor and superpixel map features. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a commonly used clinical radiological imaging technique. Due to its advantages such as being non-invasive, ionizing, and providing high contrast in soft tissue imaging, it has become the preferred imaging method for diagnosing brain tumors. To improve the accuracy and efficiency of brain MRI diagnosis, deep learning-based medical image processing methods are increasingly being applied to the diagnosis of brain tumors via MRI.

[0003] Early methods primarily employed statistical and machine learning approaches. 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 and imbalance of datasets significantly impact diagnostic performance. Medical image datasets typically exhibit imbalanced sample ratios, and the acquired images often contain noise, artifacts, and other interfering information, making it difficult to fully extract key lesion features. Different network layers often possess different spatial scales, leading to representational differences between features. Directly fusing lesion features from different layers can result in them being overwhelmed by redundant information, affecting the accuracy of lesion detection and classification. Furthermore, network structures of different depths exhibit significant differences in capturing image features, and the model struggles to dynamically adjust its emphasis on overall and local features. Therefore, existing deep learning-based medical image classification models cannot guarantee the accuracy and efficiency of diagnosing brain MRI tumors.

[0004] Patent CN202010962134.7 discloses a system and method for recognizing severe tumor images. Based on convolutional neural networks, it employs a two-stage, multi-model approach to automatically diagnose brain tumors using magnetic resonance imaging. By mining information, extracting features, and learning from historical diagnostic big data, it achieves high-accuracy computer-aided brain image data analysis, making diagnostic accuracy independent of physician experience and accelerating the diagnostic process. However, the key features and underlying details of lesions are difficult to balance effectively during the convolution process, resulting in slightly insufficient accuracy in tumor diagnosis.

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

[0006] To address the aforementioned technical problems, this invention provides a brain MRI image recognition system and method that integrates tensor and superpixel map features, thereby resolving the issues in the prior art. To achieve the above-mentioned objective, the technical solution adopted by this invention is as follows:

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

[0008] Image acquisition module: Acquires publicly available medical images of brain tumors and divides them proportionally for model training;

[0009] Image preprocessing module: Crops all brain MRI images acquired in the image acquisition module to the same size and performs standardization processing;

[0010] The module for improving model diagnostic efficiency uses three-dimensional tensors to capture multidimensional semantic features of lesions in brain MRI images processed by the image preprocessing module, and reduces the number of model parameters through depthwise separable convolution.

[0011] The brain MRI multi-layer semantic information module is integrated. Semantic features at different levels are grouped and spliced ​​in an orderly and proportional manner through dense connections. Multi-scale dilated convolution is used to fuse low-level detailed features and deep semantic information.

[0012] The module for mining local correlation features in brain MRI: It divides the approximate regions of brain MRI to form superpixel maps, and uses graph convolutional neural networks to deeply mine the local correlation features of lesions;

[0013] Model training and lesion segmentation module: An experimental platform was built to train BTNet-TS, and the trained model was tested on the brain MRI of the brain to be diagnosed.

[0014] A brain MRI image recognition method fusing tensor and superpixel map features, applied to a brain MRI image recognition system fusing tensor and superpixel map features, includes the following steps:

[0015] S0: Obtain publicly available medical images of brain tumors and divide them proportionally for model training;

[0016] S1: Crops all brain MRI images acquired in the image acquisition module to the same size and performs standardization processing;

[0017] S2: The processed brain MRI is captured using three-dimensional tensors to capture multidimensional semantic features of lesions, and the number of model parameters is reduced by depthwise separable convolution.

[0018] S3: Semantic features at different levels are grouped and spliced ​​in an equal order through dense connections, and multi-scale dilated convolution is used to fuse low-level detailed features and deep semantic information;

[0019] S4: Divide the approximate region of brain MRI to form a superpixel map, and use graph convolutional neural network to deeply mine the local correlation features of lesions;

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

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

[0022] S20: Constructing a 3D tensor to extract multidimensional semantic features: Define 3 sets for each residual block. Tensor graph of size And based on the size of the residual block input feature Its size is adjusted using bilinear interpolation, and the calculation formula is Equation (1):

[0023]

[0024] S21: Fusion of tensor features to enhance key semantic information: The original convolution operation is replaced with depthwise separable convolution. The input feature map is then calculated according to the network structure of the residual block. The three-dimensional tensor map is multiplied by the input feature map respectively, and then added to the output of the stacked layer according to the skip connection method in the residual block. The calculation formula adopts equations (2), (3) and (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 layers reduce the dimensionality of the intermediate layers, and the calculation formulas are (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 [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 based on the input block, the feature channels of the input and output are used to determine the tensor usage type. 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 the element superposition.

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

[0030] S31: Multi-level feature grouping: The i LR-3DA output feature maps to be fused are grouped into LF. i Adjust the size using pooling layers to make their dimensions uniform. Furthermore, the number of feature channels remains unchanged; then, the feature maps are divided equally according to the number of groups to be fused, and each feature map will be divided into sections containing i / C. i Block feature map F i Finally, F i Different blocks are sequentially spliced ​​together to form i groups of preliminarily fused feature blocks Fa. i The calculation formulas use equations (8) and (9):

[0031]

[0032] S32: Multi-scale feature extraction: Feature maps that initially fuse multi-level semantic information. i Fb is obtained by performing multi-scale dilated convolution calculation and feature fusion again. i The calculation formula is Equation (10):

[0033]

[0034] Where pad represents the spacing parameter;

[0035] S33: Group Feature Fusion: The pre-fused feature maps Fb... i After being concatenated sequentially, the output feature map with the same size as the i-th input feature map is obtained through convolution calculation. The calculation formula is Equation (11):

[0036]

[0037] The input feature map LF consists of multiple sets of feature maps of different sizes, and the output feature map Out has a shape of [B,C,M,N]. Based on the input features, multi-level feature grouping is performed first, and then feature blocks are extracted from each group and spliced ​​in order. 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 in brain MRI, applied to the aforementioned module for mining local correlation features in brain MRI, includes the following steps:

[0039] S41: Constructing a superpixel map based on brain MRI: Based on the input i-th layer brain MRI feature map, the superpixel map is constructed using... Convolution merges channels and performs superpixel segmentation to obtain a superpixel map SF containing multiple segmented regions. i Then, the segmented regions of the superpixel image are used as nodes of the graph, and the adjacency relationships between regions are used as node associations, using the adjacency matrix A.i The calculation formulas are expressed as equations (12) and (13), where a jk This represents the connection between adjacent nodes; a value of 1 indicates a connection if they are connected, and a value of 0 indicates a connection otherwise.

[0040]

[0041] S42: Select node features and standardize them: Select each region a from the segmentation results. j The center coordinates, average gray value (mean), and total number of pixels (count) are used as node features; the features f in each dimension are analyzed using equation (14). j Standardization process yields z j Where μ represents the mean of each feature dimension, and σ represents the variance:

[0042]

[0043] S43: Design graph convolution extraction and superpixel feature fusion: The superpixel image and its node features are used as input to a multi-layer GCN for computation to obtain the image label F corresponding to the output channel size of the LR-3DA layer. i And perform element-wise operations to combine it with the corresponding LR-3DA layer's output feature map LF. i Multiplication to preserve global features GF i This information is passed to the next GCF-2S and guides the final tumor category diagnosis; the calculation formulas are Equations (15) and (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 lightweight MobileNet and the latest RDNet and other best models, it has improved in both Acc and F1. Furthermore, BTNet-TS can better balance stability and efficiency, proving its effectiveness in brain MRI image recognition. Attached Figure Description

[0047] Figure 1 This is a flowchart 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 for lightweight residual blocks that incorporate 3D tensor attention;

[0050] Figure 4 A diagram illustrating a dense connection fusion strategy for multi-scale feature grouping;

[0051] Figure 5 This is a graph convolution feature extraction map based on superpixel segmentation;

[0052] Figure 6 This is a confusion matrix of tumor diagnosis results from multiple models on the Brain Tumor MRIDataset dataset.

[0053] Figure 7 The confusion matrix of tumor diagnosis results from multiple models on the Brain Tumor Dataset dataset;

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

[0055] Figure 9 This is a graph showing the change in loss during model training. Detailed Implementation

[0056] The following will be based on embodiments of the present invention. Figures 1-9 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

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

[0058] Image acquisition module: Acquires publicly available medical images of brain tumors and divides them proportionally for model training;

[0059] Image preprocessing module: Crops all brain MRI images acquired in the image acquisition module to the same size and performs standardization processing;

[0060] The module for improving model diagnostic efficiency uses three-dimensional tensors to capture multidimensional semantic features of lesions in brain MRI images processed by the image preprocessing module, and reduces the number of model parameters through depthwise separable convolution.

[0061] The brain MRI multi-layer semantic information module is integrated. Semantic features at different levels are grouped and spliced ​​in an orderly and proportional manner through dense connections. Multi-scale dilated convolution is used to fuse low-level detailed features and deep semantic information.

[0062] The module for mining local correlation features in brain MRI: It divides the approximate regions of brain MRI to form superpixel maps, and uses graph convolutional neural networks to deeply mine the local correlation features of lesions;

[0063] Model training and lesion segmentation module: An experimental platform was built to train BTNet-TS, and the trained model was tested on the brain MRI of the brain to be diagnosed.

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

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

[0066] Step S0: Image Acquisition: Acquire publicly available medical images of brain tumors and divide them proportionally for model training.

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

[0068] Step S2: Improve model diagnostic efficiency: Use three-dimensional tensors to capture multidimensional semantic features of lesions from the preprocessed brain MRI in step S1, and effectively reduce the number of model parameters through depthwise separable convolution.

[0069] Step S3: Fusion of multi-layer semantic information from brain MRI: Semantic features at different levels in S2 are grouped and spliced ​​in an equal order through dense connections, and multi-scale dilated convolution is used to fuse low-level detail features and deep semantic information.

[0070] Step S4: Mining Local Correlation Features in Brain MRI: The approximate regions of brain MRI in S2 are divided into superpixel maps, and the local correlation features of lesions are deeply mined based on graph convolutional neural networks.

[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 of the brain 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 a multi-scale feature fusion strategy by GroupDenseNet (MFS-GD). Figure 2 As shown.

[0073] Step S2, a method to improve model efficiency, such as... Figure 3 As shown, the process of improving the model's diagnostic efficiency is as follows:

[0074] S20: Constructing a 3D tensor to extract multidimensional semantic features. To efficiently extract key features from the input feature map, three sets of... Tensor graph of size And based on the size of the residual block input feature Its size is adjusted using bilinear interpolation. The calculation process is shown in equation (1).

[0075]

[0076] S21: Fusion of Tensor Features to Enhance Key Semantic Information. To improve the computational efficiency of residual mapping, the original convolution operation is replaced with depthwise separable convolution (DSC), and the input feature map is then computed according to the network structure of the residual blocks. Simultaneously, the 3D tensor map is multiplied by the input feature map and then added to the output of the stacked layers according to the skip connections within the residual blocks. Specifically, the basic LR-3DA structure is as follows: Figure 3 As shown, it mainly adopts The convolution is calculated as shown in equations (2), (3), and (4), where RB represents batch normalization and the ReLU activation function.

[0077]

[0078]

[0079] For deep networks, a three-layer bottleneck block replaces the original residual block, resulting in an enhanced LR-3DA structure, as shown below. Figure 3 As shown. It is achieved by introducing The convolutional layers reduce the dimensionality of the intermediate layers, which not only reduces the number of parameters but also helps the network learn more efficient feature representations, effectively alleviating the problems of gradient vanishing and gradient exploding. Their calculation methods are 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 [B, C, M, N]. 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 tensor usage type is determined using the feature channels of the input and output. After calculation by the basic residual block (BR) and the enhanced residual block (SR), the feature map Out of the residual block is output according to the element-wise superposition.

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

[0083] S31: Multi-level feature grouping. First, the i LR-3DA output feature maps LF to be fused are... i Adjust the size using pooling layers to make their dimensions uniform. Furthermore, the number of feature channels remains unchanged. Then, the feature maps are divided evenly according to the number of groups to be fused, and each feature map will be divided into sections containing i / C. i Block feature map F i Finally, F i Different blocks are sequentially spliced ​​together to form i groups of preliminarily fused feature blocks Fa. i The process is shown in equations (8) and (9).

[0084]

[0085]

[0086] S32: Multi-scale feature extraction. To capture information at different scales in the fused feature map and improve the model's robustness and expressive power, the feature map Fa, which initially fuses multi-level semantic information, is used. i Fb is obtained by calculating through multi-scale dilated convolution (DConv) and then fusing features again. i The specific process is shown in equation (10).

[0087]

[0088] Here, pad represents the spacing parameter, which determines the distance between sampling points in the convolution kernel. By adjusting this parameter value, the receptive field of the convolution kernel can be expanded, thereby capturing a wider range of spatial information.

[0089] S33: Feature Fusion by Grouping. Although concatenating multiple feature maps provides richer information, further mining of high-level features is still needed to improve the network's expressive and generalization capabilities. Therefore, the initially fused feature map Fb... i After being concatenated in sequence, the input feature maps are convolved again to obtain an output feature map with the same size as the i-th input feature map. The process is shown in Equation (11).

[0090]

[0091] The input feature map LF consists of multiple sets of feature maps of different sizes, and the output feature map Out has a shape of [B, C, M, N]. Based on the input features, multi-level feature grouping is performed, and feature blocks are extracted from each group and concatenated sequentially. 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.

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

[0093] S41: Constructing a superpixel map based on brain MRI. This is done by using 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 regions. i (a1,a2,a3,…,a j Next, the segmented regions of the superpixel image are used as nodes of the graph, and the adjacency relationships between regions are used as node associations, using the adjacency matrix A. i The calculation process is shown in equations (12) and (13), where a jk This represents the connection between adjacent nodes; it is 1 if they are connected and 0 otherwise.

[0094]

[0095]

[0096] S42: Selecting and Standardizing Node Features. Although node information in the image is obtained through superpixel segmentation, corresponding node features still need to be selected for GCN calculation. Therefore, this paper selects each region a from the segmentation results. jThe center coordinates (x, y), the average gray value (mean), and the total number of pixels (count) are used as node features. Due to the heterogeneity among node features, features f in each dimension are processed according to equation (14). j Standardization process yields z j Where μ represents the mean of each feature dimension, and σ represents the variance.

[0097]

[0098] S43: Design graph convolution extraction and superpixel feature fusion. The superpixel image and its node features are used as input to a multi-layer GCN for computation, obtaining the image label F corresponding to the output channel size of the LR-3DA layer. i And perform element-wise operations to combine it with the corresponding LR-3DA layer's output feature map LF. i Multiplication to preserve global features GF i It is passed to the next GCF-2S and guides the final tumor category 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 preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, alterations, or substitutions made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A brain MRI image recognition method that integrates tensor and superpixel map features, comprising the following steps: S0: Obtain publicly available medical images of brain tumors and divide them proportionally for model training; S1: Crops all brain MRI images acquired in the image acquisition module to the same size and performs standardization processing; S2: The processed brain MRI is captured using three-dimensional tensors to capture multidimensional semantic features of lesions, and the number of model parameters is reduced by depthwise separable convolution. S3: Semantic features at different levels are grouped and spliced ​​in an equal order through dense connections, and multi-scale dilated convolution is used to fuse low-level detailed features and deep semantic information; S4: The brain MRI region is divided into superpixel maps, and the local correlation features of the lesions are deeply mined based on graph convolutional neural networks; S5: Build an experimental platform to train BTNet-TS and test the trained model on the brain MRI of the patient to be diagnosed. S2 includes: S20: Constructing a 3D tensor to extract multidimensional semantic features: Define 3 sets of 7x7 tensor graphs for each residual block. P 7x7 And based on the size of the residual block input feature C x M x N Its size is adjusted using bilinear interpolation, and the calculation formula is Equation (1): S21: Fusion of Tensor Features to Enhance Key Semantic Information: The original convolution operation is replaced with depthwise separable convolution (DSC), and the input feature map is used for this purpose. Input Then, calculations are performed according to the network structure of the residual blocks, and the 3D tensor graph is compared with the input features. Input After multiplication, the residual is added to the output of the stacked layer according to the skip connection method within the residual block. The calculation formulas are (2), (3), and (4), where RB represents batch normalization and the ReLU activation function: For deep networks, a three-layer bottleneck block is used to replace the original residual block, and a 1x1 convolutional layer is introduced to reduce the dimensionality of the intermediate layers. The calculation formulas are (5), (6) and (7): S22: Setting up the residual tensor selection mechanism: Input feature map X and the output feature map Out The shapes are all [B, C, M, N], and the three-dimensional feature tensors are... 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 tensor usage type is determined using the input and output feature channels. After calculations by the basic residual block and the enhanced residual block, the feature map of the residual block is output according to the element-wise superposition. Out .

2. The brain MRI image recognition method fusing tensor and superpixel map features according to claim 1, characterized in that, Step S3 includes the following steps: S31: Multi-level feature grouping: grouping features to be fused i Feature map of LR-3DA output LF i Adjust the size using pooling layers to make their dimensions uniform. H i x W i The number of feature channels remains unchanged; then, the feature maps are divided equally according to the number of groups to be fused, and each feature map will be divided into groups containing... i / C i Block feature map F i Finally, F i Different blocks are pieced together in sequence to form i Feature blocks of preliminary fusion Fa i The calculation formulas use equations (8) and (9): S32: Multi-scale feature extraction: Feature blocks that initially fuse multi-level semantic information Fa i By performing multi-scale dilated convolution calculations and then fusing features again, we obtain... Fb i The calculation formula is Equation (10): in, pad Indicates the spacing parameter; S33: Group Feature Fusion: This involves fusing the preliminarily fused feature maps... Fb i After being concatenated in sequence, the result is obtained by convolution calculation with the first convolution. i The output feature maps of the same input feature map size are calculated using formula (11): Among them, feature map LF i The output feature map consists of multiple sets of feature maps of different sizes. Out The shape is [B, C, M, N]. Based on the input features, multi-level feature grouping is performed first, then feature blocks are extracted from each group and concatenated in order. 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 .

3. The brain MRI image recognition method fusing tensor and superpixel map features according to claim 1, characterized in that, Step S4 includes the following steps: S41: Constructing a superpixel map based on brain MRI: Based on the input of the first... i The brain MRI feature map was processed using 1x1 convolution with merged channels and then superpixel segmented to obtain a superpixel map containing multiple segmented regions. SF i Then, the segmented regions of the superpixel image are used as nodes of the graph, and the adjacency relationships between regions are used as node associations, using an adjacency matrix. A i The calculation formulas are expressed as equations (12) and (13), where... a jk This represents the connection between adjacent nodes; a value of 1 indicates a connection if they are connected, and a value of 0 indicates a connection otherwise. S42: Select node features and standardize the process: Select each region from the segmentation results. a j Center coordinates, average gray value mean and total number of pixels count As node features; through equation (14) for each dimension feature f j Standardization process is required to obtain z j ;in Represents the mean of each feature dimension. Then, the variance is represented as follows: S43: Design Graph Convolution Extraction and Superpixel Feature Fusion: The superpixel image and its node features are used as input to a multi-layer GCN for computation, obtaining graph labels of the corresponding lightweight residual block LR-3DA with fused 3D tensor attention, and the output channel size. F i The element-wise operations are performed to combine the output feature map of the corresponding lightweight residual block LR-3DA with fused 3D tensor attention. LF i Multiply to preserve global features GF i The graph convolution feature extraction GCF-2S based on superpixel segmentation is passed to the next superpixel segmentation-based graph convolution feature extraction and guides the final tumor category diagnosis; the calculation formula adopts equations (15) and (16):

4. A brain MRI image recognition system fusing tensor and superpixel map features, applied to the brain MRI image recognition method fusing tensor and superpixel map features as described in claim 1, characterized in that, include: Image acquisition module: Acquires publicly available medical images of brain tumors and divides them proportionally for model training; Image preprocessing module: Crops all brain MRI images acquired in the image acquisition module to the same size and performs standardization processing; The module for improving model diagnostic efficiency uses three-dimensional tensors to capture multidimensional semantic features of lesions in brain MRI images processed by the image preprocessing module, and reduces the number of model parameters through depthwise separable convolution. The brain MRI multi-layer semantic information module is integrated. Semantic features at different levels are grouped and spliced ​​in an orderly and proportional manner through dense connections. Multi-scale dilated convolution is used to fuse low-level detailed features and deep semantic information. The module for mining local correlation features in brain MRI: It divides the brain MRI region into superpixel maps and uses graph convolutional neural networks to deeply mine the local correlation features of lesions; Model training and lesion segmentation module: An experimental platform was built to train BTNet-TS, and the trained model was tested on the brain MRI of the brain to be diagnosed.

Images