Glioma malignancy grading method based on ResNet-GCN network model

Abstract

The application discloses a brain glioma malignancy grading method based on a ResNet-GCN network model, and comprises the following steps: S1, data set construction and data preprocessing, acquiring an MRI image to construct a data set, and converting the data to a required format, and preprocessing the data; S2, constructing a ResNet-GCN network model, taking a ResNet model as a basis, modifying the output of a feature extractor, obtaining a feature vector set at different levels, and connecting an adjacency matrix to a GCN model as input after the adjacency matrix is constructed by using the feature vector set; S3, setting model training parameters, specifying a loss function form used, and training the ResNet-GCN network model; S4, evaluating the ResNet-GCN network model, and finally using the trained model for actual brain glioma malignancy grading. The ResNet-GCN model provided by the application can comprehensively analyze feature vectors at different levels, has better classification performance, and realizes better auxiliary diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to a method for classifying the malignancy of gliomas based on the ResNet-GCN network model. Background Technology

[0002] Gliomas are cancerous glial cells originating from the brain and spinal cord, accounting for 30% of all brain and central nervous system tumors and 80% of all malignant brain tumors. They are the most common and deadliest primary malignant intracranial tumors. Because the clinical treatment and prognostic evaluation of gliomas are closely related to their malignancy grade, the malignancy grade of gliomas is of great significance.

[0003] Currently, preoperative grading of gliomas is primarily based on the patient's clinical presentation and MRI imaging characteristics. Experienced radiologists can identify tumors relatively accurately using MRI, but for younger radiologists or those without relevant systematic training, the rates of misdiagnosis and missed diagnosis are relatively high. Furthermore, with the advent of the big data era, radiologists often face a large volume of MRI images, and this high workload can easily lead to fatigue and reduced diagnostic accuracy. Therefore, freeing radiologists from traditional image interpretation has become a significant concern.

[0004] Currently, utilizing artificial intelligence and deep learning technologies to simulate human thinking and efficiently mine image information from large datasets of medical imaging data to assist doctors in accurate identification and diagnosis has become a promising technological approach. In recent years, researchers have successively designed models for glioma malignancy grading from the perspectives of 2D and 3D images, unimodal and multimodal MRI images, multimodal MRI and CT images, and multimodel fusion of various CNN variants, achieving some success. However, most of the models proposed in these studies rely on the highest-level image features extracted for glioma grading and diagnosis, failing to effectively utilize the contextual logical relationships between image features at different levels. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a method for classifying the malignancy of gliomas based on the ResNet-GCN network model. This method derives the classification of the malignancy of gliomas by analyzing the contextual logical relationships between image features at different levels.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for grading the malignancy of gliomas based on the ResNet-GCN network model includes the following steps:

[0008] S1. Dataset Construction and Data Preprocessing: Obtain MRI images to construct a dataset, convert the data to the required format, and preprocess the data.

[0009] S2. Construct a ResNet-GCN network model. Based on the ResNet model, modify the output of the feature extractor to obtain feature vector sets at different levels. After constructing the adjacency matrix using the feature vector sets, connect the adjacency matrix as input to the GCN model.

[0010] S3. Set the model training parameters, specify the form of the loss function to be used, and train the ResNet-GCN network model.

[0011] S4. Evaluate the ResNet-GCN network model and finally use the trained model for the actual grading of the malignancy of gliomas.

[0012] Furthermore, each data sample in the dataset is a spatially continuous distribution sequence of a set of 2D slice images.

[0013] Furthermore, in step S1, the data preprocessing specifically involves:

[0014] S11. Resample MRI data samples of different slice sizes to ensure that all data samples maintain a uniform resolution and voxel spacing.

[0015] S12. Background noise is filtered out for MRI data samples acquired by different devices to remove low grayscale noise in the background area outside the brain in the data samples.

[0016] S13. Perform bias field correction on MRI data samples acquired by different devices to smooth out large differences in image grayscale values ​​within the same brain tissue region in some data samples.

[0017] S14. Perform skull dissection on MRI data samples acquired from different devices so that the final imaging state of all data samples contains only brain tissue imaging images.

[0018] Furthermore, step S12 includes:

[0019] OSTU thresholding, morphological operations, and hole filling are performed as follows:

[0020] The maximum inter-class variance threshold for a single sample is calculated using the OTU algorithm, and coarse segmentation of the brain and background is completed. Opening and closing operations of morphology are used to break the weak adhesion between the brain and background and fill any small holes that may exist inside the brain and background.

[0021] For the mask image after morphological operations, we add statistics on the number of contours and the number of contour boundary points, and filter out small holes that may exist after morphological operations by setting a threshold.

[0022] Furthermore, the ResNet-GCN network model is as follows:

[0023] Based on the ResNet18 network model, the model structure consists of 5 layers:

[0024] The first layer includes a 7×7 convolution operation, batch normalization, ReLU activation, and max pooling operation;

[0025] Layers 2 through 5 each consist of two residual blocks, and each residual block consists of two repeated operation sequences. One operation sequence is a 3×3 convolution operation, batch normalization, and ReLU activation.

[0026] For residual blocks connected across layers, an additional 1×1 convolution operation and batch normalization are added for downsampling.

[0027] GCN consists of two graph convolutional network layers and one fully connected layer. The first graph convolutional network layer has an input dimension of 512 and an output dimension of 128. The second graph convolutional network layer has an input dimension of 128 and an output dimension of 64. The fully connected layer has an input dimension of 64 and an output dimension of 3.

[0028] Furthermore, from the first to the fifth layer of the ResNet18 network, the feature vectors output by each layer are extracted, and downsampling is used to make the dimension of the feature vectors output by each layer the same as the dimension of the feature vectors of the highest level, thus obtaining the feature vector set of different levels in the ResNet18 network; the feature vector set is regarded as a set of nodes with hierarchical logical relationship and passed as input parameters to GCN; assuming that the weights of each node follow a uniform distribution, the initialization is completed.

[0029] Furthermore, an adjacency matrix is ​​constructed based on the feature matrix and used as input to two graph convolutional network layers, specifically:

[0030] The 512×5 feature vector set obtained from the ResNet18 network is regarded as the feature matrix X;

[0031] Multiplying X by its transpose yields a 5×5 adjacency matrix with autocorrelation properties. The elements of this adjacency matrix are then sorted in descending order in one dimension. The eigenvalues ​​at the top 20% of positions are selected as thresholds, and the eigenvalues ​​in the adjacency matrix are binarized. Finally, the 5×5 unit diagonal matrix is ​​superimposed on the binarized adjacency matrix, and Laplacian regularization is performed to obtain a 5×5 Laplacian matrix. The 128×5 feature matrix output from the first graph convolutional network layer is used to construct the adjacency matrix, which serves as the input to the second graph convolutional network layer. For the 64×5 feature matrix output from the second graph convolutional network layer, a 64-dimensional weighted feature vector is calculated through weighted summation. A fully connected layer is then used to obtain the final classification result.

[0032] Furthermore, step S3 specifically includes:

[0033] S31. Set a global random number seed, divide the preprocessed dataset into k folds, take k-1 folds as the training set each time, and take the remaining 1 fold as the validation set.

[0034] S32. Parameter initialization is performed using the Kaiming method, and parameter optimization is performed using the SGD optimizer.

[0035] S33. Set the batch size of training samples per session, the total number of training epochs, the initial learning rate, and the learning rate decay coefficient.

[0036] During the training of the S34 and ResNet-GCN network models, the cross-entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error backpropagation algorithm.

[0037] Furthermore, in step S3, the loss function specifically includes two parts:

[0038] ResNet uses the cross-entropy loss between the hierarchical results generated by the highest-level feature vectors and the reference hierarchical results, i.e., the first-level diagnostic loss.

[0039] GCN is based on the cross-entropy loss between the hierarchical results generated by feature vector sets at different levels and the reference hierarchical results, i.e., the second-level diagnostic loss.

[0040] The loss function of the ResNet-GCN network model is ultimately expressed as:

[0041] L=αL CE1 +βL CE2

[0042] Among them, L CE1 For Level 1 diagnostic loss, L CE2 Let α represent the second-order diagnostic loss, and β represent the weighting coefficients for the first-order and second-order diagnostic losses, respectively.

[0043] Furthermore, the ResNet-GCN network model is evaluated using precision, recall, accuracy, and F1-score as performance metrics. The meanings and calculation formulas for each metric are as follows:

[0044] Precision represents the proportion of data that are actually positive out of all data that were initially classified as positive.

[0045]

[0046] Recall rate represents the proportion of data that is actually positive but is judged as positive.

[0047]

[0048] Precision rate represents the proportion of data points that are correctly classified out of all available data.

[0049]

[0050] F1-Score, the harmonic mean of precision and recall, is a secondary evaluation indicator calculated by combining primary evaluation indicators.

[0051]

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] 1. The ResNet-GCN network model proposed in this invention can comprehensively consider image features at different levels to obtain diagnostic results with higher accuracy and greater reference value. Among them, ResNet is used as a feature extractor to extract image features at different levels of input. In practice, it can be easily replaced by other feature extractors according to specific needs, which is highly flexible. GCN is used as a classifier to derive diagnostic results based on image features at different levels. It can effectively improve the model's auxiliary diagnostic performance with a simple structure and fewer parameters.

[0054] 2. The linear weighted loss function used in this invention takes into account both the losses that may be caused by traditional classification results and the losses that may be caused by GCN classification results, which can improve the stability of the model's diagnostic results to a certain extent. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a flowchart of the data preprocessing method of the present invention;

[0057] Figure 3 This is a network model diagram of the method of the present invention;

[0058] Figure 4a This is a schematic diagram showing the changes in the Accuracy performance index of the ResNet-GCN model and the VGG16 model during training.

[0059] Figure 4b This is a schematic diagram showing the changes in the Accuracy performance index of the ResNet-GCN model and the ResNet18 model during training.

[0060] Figure 4c This is a schematic diagram showing the changes in the Accuracy performance index of the ResNet-GCN model and the ResNet50 model during training.

[0061] Figure 4d This is a schematic diagram showing the changes in the Accuracy performance index of the ResNet-GCN model and the ResNeXt50 model during training.

[0062] Figure 4e This is a schematic diagram showing the changes in the accuracy performance index of the ResNet-GCN model and the ConvNeXt-Tiny model during training. Detailed Implementation

[0063] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0064] Example

[0065] like Figure 1 As shown, the present invention provides a method for grading the malignancy of gliomas based on the ResNet-GCN network model, comprising the following steps:

[0066] S1. Dataset Construction and Data Preprocessing: Obtain MRI images to construct a dataset, convert the data to the required format, and preprocess the data; each data sample in the dataset is a spatially continuous distribution sequence of a set of 2D slice images.

[0067] In this embodiment, the internationally publicly available BraTS-2018 dataset and IXI dataset are used to train and validate the proposed ResNet-GCN network model.

[0068] The BraTS-2018 dataset contains MRI data from 285 glioma cases, including 210 high-grade gliomas and 75 low-grade gliomas. Each case has image data in four modalities: T1, T2, FLAIR, and T1ce, each corresponding to a case-level glioma malignancy label. All samples in BraTS-2018 have undergone basic preprocessing by physicians, and the images for each modality are 155×240×240 2D slice spatial sequences.

[0069] The IXI dataset provides MRI data from 578 healthy individuals, including T1, T2, and PD-weighted images. The samples in the IXI dataset come from different hospitals and different models of MR imaging equipment, and are raw MRI data without any preprocessing. The image specifications for different modalities of different cases vary significantly: for example, T1 modality slices are 256 in both the x and z axes, and vary from 100 to 150 in the y axis; T2 modality slices are 256 in both the y and z axes, and vary from 100 to 150 in the x axis.

[0070] In this embodiment, as Figure 2 As shown, the specific preprocessing of the data is as follows:

[0071] S11. Resample MRI data samples of different slice sizes to ensure that all data samples maintain a uniform resolution and voxel spacing.

[0072] S12. Filter background noise from MRI data samples acquired by different devices, removing low-grayscale noise from the external brain background region of the data samples; including:

[0073] OSTU thresholding, morphological operations, and hole filling are performed as follows:

[0074] The maximum inter-class variance threshold for a single sample is calculated using the OTU algorithm, and coarse segmentation of the brain and background is completed. Opening and closing operations of morphology are used to break the weak adhesion between the brain and background and fill any small holes that may exist inside the brain and background.

[0075] For the mask image after morphological operations, we add statistics on the number of contours and the number of contour boundary points, and filter out small holes that may exist after morphological operations by setting a threshold.

[0076] S13. Perform bias field correction on MRI data samples acquired by different devices to smooth out large differences in image grayscale values ​​within the same brain tissue region in some data samples.

[0077] S14. Perform skull dissection on MRI data samples acquired from different devices so that the final imaging state of all data samples contains only brain tissue imaging images.

[0078] In this embodiment, to use sample data from both datasets simultaneously, preprocessing is required to standardize the image specifications of all samples. Using the sample image specifications from the BraTS-2018 dataset as a benchmark, image resampling, background denoising, bias field correction, and skull removal are sequentially performed on the sample images from the IXI dataset, ultimately resulting in sample images with a resolution of 155×240×240, a voxel spacing of 1×1×1, and skull removal.

[0079] Since the total number of samples in the IXI dataset is 578, exceeding the total number of samples in the BraTS-2018 dataset, to ensure the balance of the number of samples of each type in the newly constructed dataset to the greatest extent, 210 samples of T1 modality data were randomly selected from the IXI dataset and mixed with the T1ce modality data of all samples in the BraTS-2018 dataset to obtain the BraTS-IXI dataset. Data augmentation operations such as image flipping, rotation, and brightness adjustment were performed on a portion of the data samples in the BraTS-IXI dataset, and all data samples were normalized.

[0080] S2. Construct a ResNet-GCN network model. Based on the ResNet model, modify the output of the feature extractor to obtain feature vector sets at different levels. After constructing the adjacency matrix using the feature vector sets, connect the adjacency matrix as input to the GCN (Graph Convolutional Network) model.

[0081] The ResNet-GCN network model is as follows:

[0082] Based on the ResNet18 network model, the model structure consists of 5 layers:

[0083] The first layer includes a 7×7 convolution operation, batch normalization, ReLU activation, and max pooling operation;

[0084] Layers 2 through 5 each consist of two residual blocks, and each residual block consists of two repeated operation sequences. One operation sequence is a 3×3 convolution operation, batch normalization, and ReLU activation.

[0085] For residual blocks connected across layers, an additional 1×1 convolution operation and batch normalization are added for downsampling.

[0086] GCN consists of two graph convolutional network layers and one fully connected layer. The first graph convolutional network layer has an input dimension of 512 and an output dimension of 128. The second graph convolutional network layer has an input dimension of 128 and an output dimension of 64. The fully connected layer has an input dimension of 64 and an output dimension of 3.

[0087] From the first to the fifth layer of the ResNet18 network, the feature vectors output by each layer are extracted, and downsampling is used to make the dimension of the feature vectors output by each layer the same as the dimension of the feature vectors of the highest level, thus obtaining the feature vector set of different levels in the ResNet18 network; the feature vector set is regarded as a set of nodes with hierarchical logical relationship and passed as input parameters to GCN; assuming that the weights of each node follow a uniform distribution, the initialization is completed.

[0088] An adjacency matrix is ​​constructed based on the feature matrix and used as the input to two graph convolutional network layers;

[0089] The 512×5 feature vector set obtained from the ResNet18 network is regarded as the feature matrix X;

[0090] Multiplying X by its transpose yields a 5×5 adjacency matrix with autocorrelation properties. The elements of this adjacency matrix are then sorted in descending order in one dimension. The eigenvalues ​​at the top 20% of positions are selected as thresholds, and the eigenvalues ​​in the adjacency matrix are binarized. Finally, the 5×5 unit diagonal matrix is ​​superimposed on the binarized adjacency matrix, and Laplacian regularization is performed to obtain a 5×5 Laplacian matrix. The 128×5 feature matrix output from the first graph convolutional network layer is used to construct the adjacency matrix, which serves as the input to the second graph convolutional network layer. For the 64×5 feature matrix output from the second graph convolutional network layer, a 64-dimensional weighted feature vector is calculated through weighted summation. A fully connected layer is then used to obtain the final classification result.

[0091] like Figure 3 The diagram shown is a structural diagram of the ResNet-GCN network model.

[0092] S3. Set the model training parameters, specify the form of the loss function to be used, and train the ResNet-GCN network model, specifically including:

[0093] S31. Set a global random number seed, divide the preprocessed dataset into k folds, take k-1 folds as the training set each time, and take the remaining 1 fold as the validation set.

[0094] S32. Parameter initialization is performed using the Kaiming method, and parameter optimization is performed using the SGD optimizer.

[0095] S33. Set the batch size of training samples per session, the total number of training epochs, the initial learning rate, and the learning rate decay coefficient.

[0096] In this embodiment, the random number seed is preset to 42, and the dataset is divided into three folds based on this seed. Two folds are used as the training set, and the remaining fold is used as the validation set. The weight decay index is set to 1e-5. The batch size is set to 16. The total number of epochs is set to 150. The initial learning rate is set to 0.001, which decays to 0.5 times every 50 epochs.

[0097] During the training of the S34 and ResNet-GCN network models, the cross-entropy loss function is used to calculate the loss, and the weights and biases in the model are updated through the error backpropagation algorithm.

[0098] In this embodiment, the loss function specifically includes two parts:

[0099] ResNet uses the cross-entropy loss between the hierarchical results generated by the highest-level feature vectors and the reference hierarchical results, i.e., the first-level diagnostic loss.

[0100] GCN is based on the cross-entropy loss between the hierarchical results generated by feature vector sets at different levels and the reference hierarchical results, i.e., the second-level diagnostic loss.

[0101] The loss function of the ResNet-GCN network model is ultimately expressed as:

[0102] L=αL CE1 +βL CE2

[0103] Among them, L CE1 For Level 1 diagnostic loss, L CE2 The loss is the secondary diagnostic loss, and α and β are the weighting coefficients corresponding to the primary diagnostic loss and the secondary diagnostic loss, respectively. In this embodiment, both are set to 0.5.

[0104] S4. Evaluate the ResNet-GCN network model and finally use the trained model for the actual grading of the malignancy of gliomas.

[0105] The ResNet-GCN network model is evaluated using precision, recall, accuracy, and F1-score as performance metrics. The meanings and calculation formulas for each metric are as follows:

[0106] Precision represents the proportion of data that are actually positive out of all data that were initially classified as positive.

[0107]

[0108] Recall rate represents the proportion of data that is actually positive but is judged as positive.

[0109]

[0110] Precision rate represents the proportion of data points that are correctly classified out of all available data.

[0111]

[0112] F1-Score, the harmonic mean of precision and recall, is a secondary evaluation indicator calculated by combining primary evaluation indicators.

[0113]

[0114] In this embodiment, based on the above performance metrics, the test results on the BraTS-IXI dataset are shown in Table 1 below. As can be seen from the experimental results in Table 1, compared with a series of classic architecture models, the ResNet-GCN network model proposed in this invention can comprehensively analyze feature vectors at different levels, showing a certain degree of improvement in various model evaluation metrics and exhibiting better classification performance. The changes in the Accuracy metric of each model during training are shown below. Figure 4a , Figure 4b , Figure 4c , Figure 4d as well as Figure 4e As shown.

[0115] Net-Name Accuracy (%) Precision (%) Recall (%) F1-Score VGG16 90.60 86.12 86.80 86.61 ResNet18 92.85 89.10 90.13 89.61 ResNet50 92.56 89.66 88.90 89.28 ResNeXt50 91.39 87.34 86.46 86.90 ConvNeXt-Tiny 91.38 88.21 85.37 86.77 ResNet-GCN 94.12 92.03 91.72 91.87

[0116] Table 1

[0117] Finally, the trained ResNet-GCN network model was obtained, which can be used to classify the malignancy of gliomas based on case images after data format conversion.

[0118] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0119] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for grading the malignancy of glioma based on a ResNet-GCN network model, characterized in that, Comprise the following steps: S1, data set construction and data preprocessing, obtain MRI image construction data set, and convert the data to the required format, and pre-process the data; S2, construct a ResNet-GCN network model, modify the output of the feature extractor based on the ResNet model, obtain a set of feature vectors at different levels, construct an adjacency matrix using the set of feature vectors, and connect the adjacency matrix as input to the GCN model; the ResNet-GCN network model is specifically: Based on the ResNet18 network model as the basic framework, the model structure is divided into 5 layers; From the first layer to the fifth layer of the ResNet18 network, the feature vectors output by each layer are introduced, and the dimension of the feature vectors output by each layer is made the same as the dimension of the highest level feature vector through downsampling, to obtain a set of feature vectors at different levels in the ResNet18 network; the feature vector set is regarded as a set of nodes with hierarchical logical relationship, and is input to the GCN; assuming that the weight of each node is subject to uniform distribution, initialization is completed; Based on the feature matrix, an adjacency matrix is constructed as the input of two graph convolution network layers, specifically: For the 512*5 feature vector set obtained from the ResNet18 network, it is regarded as a feature matrix X; Multiply X with its transpose matrix to obtain a 5*5 adjacency matrix with autocorrelation properties, then sort the adjacency matrix elements in descending order in one dimension, select the top 20% of the feature values as the threshold, binarize the feature values in the adjacency matrix, and finally superimpose a 5*5 unit diagonal matrix with the binarized adjacency matrix to perform Laplace regularization to obtain a 5*5 Laplace matrix; use the 128*5 feature matrix output by the first graph convolution network layer to construct an adjacency matrix as the input of the second graph convolution network layer; for the 64*5 feature matrix output by the second graph convolution network layer, calculate the weighted feature vector with a dimension of 64 by weighted summation, and then use the fully connected layer to obtain the final classification result; S3, set the model training parameters, specify the form of the loss function used, and train the ResNet-GCN network model; S4, evaluate the ResNet-GCN network model, and finally use the trained model for actual brain glioma malignancy grading.

2. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 1, characterized in that, Each data sample in the data set is a spatially continuous distribution sequence of a group of 2D slice images.

3. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 1, characterized in that, In step S1, the data preprocessing is specifically: S11, image resampling is performed on MRI data samples of different slice specifications to maintain uniform resolution and voxel spacing for all data samples; S12, background noise filtering is performed on MRI data samples collected by different devices to remove low gray value noise in the external background area of the data samples; S13, bias field correction is performed on MRI data samples collected by different devices to smooth the large image gray value difference in the same brain tissue region of the data samples; S14, skull stripping is performed on the MRI data samples collected by different devices, so that only brain tissue imaging is included in the final imaging state of all data samples.

4. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 3, characterized in that, Step S12 includes: OSTU threshold segmentation, morphological operation and hole filling, specifically: According to the OSTU algorithm, the maximum inter-class variance threshold of a single sample is calculated, and coarse segmentation of the brain and background is completed; using the opening and closing operations of morphological operation, the weak adhesion between the brain and the background is disconnected, and the small size holes that may exist inside the brain and the background are filled; For the Mask image after morphological operation, the number of contours and the number of contour boundary points are added, and small size holes that may exist after morphological operation are filtered out by setting a threshold.

5. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 1, characterized in that, The first layer of the ResNet-GCN network model includes a 7x7 convolution operation, batch normalization, ReLU activation and max pooling operation; The second to fifth layers each include two residual blocks, and each residual block includes two repeated operation sequences, one of which is a 3x3 convolution operation, batch normalization and ReLU activation; For the residual blocks connected across layers, an additional 1x1 convolution operation and batch normalization are added for down-sampling; The GCN is composed of two graph convolution network layers and a fully connected layer; the input dimension of the first graph convolution network layer is 512, and the output dimension is 128; the input dimension of the second graph convolution network layer is 128, and the output dimension is 64; the input dimension of the fully connected layer is 64, and the output dimension is 3.

6. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 1, characterized in that, Step S3 specifically includes: S31, set the global random seed, divide the preprocessed data set into k folds, and each time take k-1 folds as the training set and the remaining 1 fold as the validation set; S32, parameter initialization is performed using the Kaiming method, and parameter optimization is performed using the SGD optimizer; S33, set the single training sample batch size BatchSize, the total number of training rounds epoch, the initial learning rate and the learning rate decay coefficient; S34, the cross-entropy loss function is used to calculate the loss during the training process of the ResNet-GCN network model, and the weights and biases in the model are updated through the error back propagation algorithm.

7. The brain glioma malignancy grading method based on the ResNet-GCN network model according to claim 1, characterized in that, In step S3, the loss function specifically includes two parts: The cross-entropy loss between the hierarchical results generated based on the highest level feature vector and the reference hierarchical results, i.e., the first-level diagnosis loss; The cross-entropy loss between the hierarchical results generated based on the feature vector set of different levels and the reference hierarchical results, i.e., the second-level diagnosis loss; The loss function of the ResNet-GCN network model is finally represented as: L= α L CE1 + β L CE2 wherein L CE1 is the primary diagnostic loss, L CE2 is the secondary diagnostic loss, α and β are the weighting coefficients corresponding to the primary diagnostic loss and the secondary diagnostic loss, respectively.

8. The ResNet-GCN network model-based glioma malignancy grading method according to claim 1, characterized in that, The evaluation of the ResNet-GCN network model specifically uses precision, recall, accuracy and F1-Score as evaluation indicators of the performance of the network model, and the meanings and calculation formulas of the indicators are as follows: Precision, which represents the proportion of actual positive samples in all samples judged as positive samples: Recall, which represents the proportion of samples judged as positive samples in all actual positive samples: Precision, which indicates the proportion of the sample data that is correctly determined in all sample data: F1-Score, the harmonic mean of precision and recall, is a secondary evaluation index calculated from the primary evaluation index: 。

Images