A method and apparatus for classifying brain tumor magnetic resonance images.
By constructing a classification model that integrates channel attention mechanism, multi-scale convolutional neural network and support vector machine, the problem of low classification accuracy of brain tumor magnetic resonance images was solved, and more efficient and stable image classification was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANGGANG NORMAL UNIV
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-26
AI Technical Summary
Existing methods for classifying brain tumor MRI images suffer from insufficient feature extraction, resulting in low classification accuracy, and the stability and computational complexity of the models need to be optimized.
A classification model integrating channel attention mechanism, multi-scale convolutional neural network and support vector machine is constructed. Image features are extracted by multi-scale convolutional kernels and feature fusion is performed by combining channel attention mechanism and support vector machine. The training process is optimized by hot restart stochastic gradient descent algorithm.
It improves the classification accuracy and model stability of brain tumor magnetic resonance images, reduces computational complexity, and enhances the ability to perceive multi-level information in brain tumor images and the robustness of classification.
Smart Images
Figure CN122090130A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image classification technology, and more specifically, relates to a method and apparatus for classifying magnetic resonance images of brain tumors. Background Technology
[0002] Traditional medical image analysis methods rely on manual examination, which is inefficient and easily affected by subjective factors. With the development of deep learning, convolutional neural networks (CNNs) have achieved remarkable results in medical image analysis.
[0003] However, when processing magnetic resonance imaging (MRI) images of brain tumors, existing methods still suffer from insufficient feature extraction, leading to low classification accuracy. Furthermore, the stability and computational complexity of the classification models require further optimization and improvement. Summary of the Invention
[0004] This invention provides a method and apparatus for classifying brain tumor magnetic resonance images, thereby solving the problem of low accuracy in the classification of brain tumor magnetic resonance images in the prior art.
[0005] This invention provides a method for classifying magnetic resonance images of brain tumors, comprising the following steps: Acquire a dataset of magnetic resonance imaging (MRI) images of brain tumors; A classification model is constructed that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The constructed classification model is trained using the dataset to obtain a trained classification model; The brain tumor MRI image to be classified is input into the trained classification model, and the classification result is output.
[0006] Preferably, the images in the dataset are preprocessed and labeled with brain tumor categories.
[0007] Preferably, the preprocessing includes at least one of normalization and data augmentation; the normalization includes at least one of size normalization and grayscale normalization; the data augmentation includes at least one of random rotation, horizontal / vertical flipping, and random scaling.
[0008] Preferably, the multi-scale convolutional neural network uses convolutional kernels of various sizes to extract image features in parallel.
[0009] Preferably, the multi-scale convolutional neural network includes three parallel convolutional paths, each of which consists of a convolutional layer, a batch normalization layer, and a max pooling layer connected in sequence; the convolutional kernels of the three convolutional paths are 3×3, 5×5, and 7×7, respectively.
[0010] Preferably, the training process of the classification model is optimized using the hot-restart stochastic gradient descent algorithm.
[0011] Preferably, the hot-restart stochastic gradient descent algorithm combines a stochastic gradient descent optimizer with a hot-restart mechanism; cosine annealing with hot restart is used as the learning rate scheduler to assist the stochastic gradient descent optimizer in training the classification model.
[0012] Preferably, the classification method for brain tumor magnetic resonance images further includes: performance evaluation of the trained classification model; wherein the performance evaluation index includes accuracy, which is the proportion of correctly predicted samples to the total number of samples.
[0013] Preferably, the performance evaluation metrics also include at least one of recall, precision, F1 score, macro average, and micro average; wherein, precision is the proportion of all samples that the model predicts to be positive samples, but which are actually positive samples.
[0014] On the other hand, the present invention provides a classification device for magnetic resonance images of brain tumors, comprising: Dataset building unit, used to acquire brain tumor magnetic resonance image datasets; The model building unit is used to build a classification model that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The model training unit is used to train the constructed classification model using the dataset to obtain the trained classification model. The image classification unit is used to input the brain tumor MRI image to be classified into the trained classification model and output the classification result. The brain tumor magnetic resonance image classification device is used to perform the steps in the brain tumor magnetic resonance image classification method described above.
[0015] One or more technical solutions provided in this invention have at least the following technical effects or advantages: (1) This invention constructs a classification model that integrates channel attention mechanism, multi-scale convolutional neural network, and support vector machine; wherein, the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier. This invention classifies brain tumor magnetic resonance images based on the above classification model (denoted as MCACNN-SVM model). Since this invention uses multi-scale convolutional neural network (MCACNN), it can effectively capture spatial features of different scales in brain tumor magnetic resonance images by using convolutional kernels of different scales, and can identify details and global information in brain tumor magnetic resonance images, effectively improving the recognition ability of tumor edges, textures and other features, thus improving the multi-level information perception ability of images. Since this invention introduces channel attention mechanism (SEnet) to improve the performance of convolutional neural network, this mechanism can enhance the network's attention to key features when recognizing brain tumors by adaptively adjusting the weights of feature channels, thus improving the classification accuracy. Furthermore, this invention achieves feature-level fusion by fusing features extracted by convolutional neural networks (CNNs) with support vector machines (SVMs). The CNN extracts deep spatial features from brain tumor MRI images, while the SVM performs classification based on these features. Leveraging SVM's superior performance under small sample conditions, the accuracy and robustness of classification are improved. Moreover, the SVM effectively optimizes classification boundaries, clearly distinguishing different tumor types when handling complex classification tasks, thus enhancing the model's classification stability. In summary, this invention not only improves the accuracy of brain tumor MRI image classification but also enhances the stability of the classification model.
[0016] (2) The present invention uses the hot restart stochastic gradient descent (SGDR) algorithm to optimize the training process of the classification model, which can accelerate the convergence of the model and improve the generalization ability. With the introduction of multi-scale convolution and channel attention mechanism, the computational complexity of the model is effectively controlled. Attached Figure Description
[0017] Figure 1 This is a framework diagram of the classification model used in the classification method for brain tumor magnetic resonance images provided in Embodiment 1 of the present invention; Figure 2 This is a confusion matrix for classifying brain tumor magnetic resonance images using a classification method for brain tumor magnetic resonance images provided in Embodiment 1 of the present invention. Figure 3 The ROC curve for classifying brain tumor magnetic resonance images using the classification method for brain tumor magnetic resonance images provided in Embodiment 1 of the present invention is shown. Detailed Implementation
[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0019] Example 1: Example 1 provides a method for classifying brain tumor magnetic resonance images, including the following steps: Acquire a dataset of magnetic resonance imaging (MRI) images of brain tumors; A classification model is constructed that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The constructed classification model is trained using the dataset to obtain a trained classification model; The brain tumor MRI image to be classified is input into the trained classification model, and the classification result is output.
[0020] Example 1 employs a multi-scale convolutional neural network (CNN), utilizing convolutional kernels of different scales to effectively capture spatial features at different scales in brain tumor MRI images. This enables the identification of details and global information within the images, significantly improving the recognition of features such as tumor edges and textures, thus enhancing the perception of multi-level information in the images. Example 1 also introduces a channel attention mechanism to improve the performance of the CNN. This mechanism adaptively adjusts the weights of feature channels, enhancing the network's focus on key features when identifying brain tumors, thereby improving classification accuracy. By explicitly modeling the dependencies between different channels and dynamically adjusting the weights of each channel, the channel attention mechanism adaptively improves the response of important feature channels and reduces interference from redundant information. Furthermore, Example 1 achieves feature-level fusion by fusing features extracted by the CNN with a support vector machine (SVM). The CNN extracts deep spatial features from the brain tumor MRI images, and the SVM performs classification based on these features, leveraging its superior performance under small sample conditions to improve classification accuracy and robustness. Furthermore, support vector machines can effectively optimize classification boundaries, clearly distinguishing different types of tumors when handling complex classification tasks, thus enhancing the model's classification stability. In other words, Example 1 not only improves the accuracy of brain tumor MRI image classification but also enhances the stability of the classification model.
[0021] The images in the dataset are preprocessed and labeled with brain tumor categories. The preprocessing includes normalization and data augmentation; normalization includes size normalization and grayscale normalization; data augmentation includes random rotation, horizontal / vertical flipping, and random scaling.
[0022] The dataset can be specifically divided into a training set, a validation set, and a test set. The model is trained using the training set. Hyperparameters are iteratively optimized using the validation set; the optimal model is automatically saved when the preset maximum number of iterations is reached. The trained model is then used to perform classification and diagnostic tasks on the test set.
[0023] See Figure 1 To overcome the limitations of traditional convolutional neural networks (CNNs) in feature extraction, this invention proposes a classification model that integrates multi-scale CNNs of different scales to enhance the perception of information at different scales in images. In the architecture of the classification model, this invention carefully adjusts and extends the traditional CNN structure to meet the specific needs of brain tumor MRI image classification. By introducing multi-scale convolutional kernels and pooling windows, the model can capture richer image features, thereby improving the network's ability to recognize edges, textures, and more complex patterns. Furthermore, a support vector machine (SVM) is used as the final classifier, responsible for optimizing the classification of high-level features extracted by the multi-scale channel attention convolutional neural network (MCACNN). The SVM constructs an optimal hyperplane to partition the decision boundary in the feature space, a method that maintains strong generalization ability when processing high-dimensional feature data. The nuanced feature extraction of MCACNN and the discriminative classification of SVM complement each other, creating a powerful framework (MCACNN-SVM) for accurate image analysis.
[0024] Specifically, the multi-scale convolutional neural network uses convolutional kernels of various sizes to extract image features in parallel. For example, the multi-scale convolutional neural network includes three parallel convolutional paths, each consisting of a convolutional layer (Conv), a batch normalization layer (BN), and a max pooling layer connected in sequence; the convolutional kernels of the three paths are 3×3, 5×5, and 7×7, respectively. The original single-scale convolutional kernels are replaced with three different sizes of convolutional kernels (3×3, 5×5, and 7×7) to capture features at different scales. After completing the convolutional operations at the three different scales, a 1×1 convolutional layer is applied to concatenate the inputs from the three scales and achieve channel compression. This design enables the model to capture more comprehensive details and structural features in brain tumor images. This parallel structure ensures that the features extracted from convolutional kernels at different scales can be learned and updated independently, thereby avoiding potential interference between multi-scale features. The outputs from multiple branches are fused through concatenation and then compressed using 1×1 convolutions, mapping the complex nonlinear features of brain tumors into a more linear feature space. While softmax layers in convolutional neural networks perform well in classifying linearly separable data, they are less effective at handling complex decision boundaries. In contrast, support vector machines (SVMs) achieve clearer feature space separation. Unlike fully connected layers using softmax, SVMs utilize the maximum margin separation principle to establish more robust decision boundaries in the high-dimensional feature space extracted by deep networks, thereby improving classification stability and generalization ability. Therefore, in this model, we first use convolutional neural networks for feature extraction, and finally use SVMs to complete the classification task, achieving fusion of the two at the feature level. By introducing multi-scale convolution and pooling operations, the improved MCACNN network can better capture multi-scale information in images. This enhances the network's ability to perceive information at different scales, thereby improving the accuracy and robustness of image classification tasks. With the help of multi-scale convolution and pooling operations, this network has a more comprehensive understanding of image features and can efficiently process information at various scales.
[0025] Considering that brain tumor MRI images typically exhibit complex textures and structural features, this invention selects a channel attention mechanism. Channel attention effectively helps the network extract and focus on important feature channels. By enhancing the weights of key features, this mechanism can both accelerate the extraction of key features from brain tumor images and suppress irrelevant features, thereby reducing computational load and significantly improving recognition and classification performance. Furthermore, compared to other types of attention mechanisms, channel attention has lower computational complexity, making its implementation and training in neural networks simpler.
[0026] Furthermore, this invention preferably employs a hot-restart stochastic gradient descent algorithm to optimize the training process of the classification model. That is, to accelerate model convergence and improve generalization ability, this invention uses a hot-restart stochastic gradient descent strategy. This strategy, by periodically adjusting the learning rate, avoids the model getting trapped in local optima and promotes rapid convergence to the global optimum.
[0027] The hot-restart stochastic gradient descent algorithm combines a stochastic gradient descent optimizer with a hot-restart mechanism; it uses cosine annealing with hot restart as a learning rate scheduler to assist the stochastic gradient descent optimizer in training the classification model.
[0028] Specifically, the core idea of the warm restart mechanism is to set a periodic loop during training. When this loop ends, the learning rate is reset to its initial value and training continues. Furthermore, after each restart, the model uses the parameters from the previous restart as its initial state. This mechanism essentially constitutes a method for dynamically adjusting the learning rate. The accuracy of the classification model will vary depending on the number of iterations under different configurations of the optimizer and learning rate scheduler. The warm restart strategy, combined with cosine annealing, adjusts the learning rate cyclically: starting from a low value, gradually increasing to a peak, and then gradually decreasing according to the cosine curve. After each annealing cycle, the warm restart mechanism resets the learning rate to its initial value and starts a new round of annealing. This strategy aims to help the model escape local optima by periodically increasing the learning rate, thereby avoiding getting trapped in local minima and improving the model's convergence performance.
[0029] Furthermore, the classification method for brain tumor magnetic resonance images may also include: performance evaluation of the trained classification model. The performance evaluation metrics primarily include accuracy, which is the proportion of correctly predicted samples out of the total number of samples. The performance evaluation metrics may also include at least one of recall, precision, F1 score, macro-average, and micro-average. Recall is the proportion of all actually positive samples correctly identified as positive. Precision is the proportion of all samples predicted as positive by the model that are actually positive. The F1 score is the harmonic mean of precision and recall, designed to balance these two metrics, and is particularly meaningful in cases of imbalanced data. Two summary metrics, macro-average and micro-average, are introduced to more comprehensively evaluate the overall performance of the classification model. Due to the imbalanced distribution of classes in the data, macro-average is specifically used as an evaluation metric for model performance. Macro-average is a simple average of the metrics for each class (such as precision, recall, and F1 score), assigning equal weight to each class without considering the differences in the number of samples in each class. Macro-means emphasize the model's balanced performance across classes and are suitable for performance comparisons when class distributions are imbalanced. Micro-means are calculated by accumulating true positives, false positives, and false negatives across all classes and then calculating relevant metrics. Micro-means assigns equal weights to all samples, thus better reflecting the model's overall classification ability across the entire dataset.
[0030] In summary, the preferred solution provided by the present invention mainly includes the following aspects: (1) preprocessing of the dataset; (2) extracting multi-level features of brain tumor magnetic resonance images using multi-scale convolution kernels; (3) introducing channel attention mechanism to optimize feature representation; (4) classifying the extracted features using support vector machine; (5) optimizing the model training process by using hot restart stochastic gradient descent; and (6) evaluating the performance of the model.
[0031] To evaluate the effectiveness of the classification model proposed in this invention, a publicly available brain tumor MRI dataset was used for validation. This dataset, also hosted on the Kaggle platform, contains 7,023 T1-weighted contrast-enhanced magnetic resonance imaging (CE-MRI) slices. These images were acquired from real clinical cases in hospitals and underwent rigorous preprocessing, including anonymization, craniotomy, and registration, to ensure data quality and consistency.33,34 The dataset was divided into four categories: glioma, meningioma, pituitary adenoma, and no tumor, thus forming a roughly balanced multi-class classification task.
[0032] In addition, to enhance the generalization ability of the model and alleviate the overfitting problem, we implemented a comprehensive data preprocessing and augmentation process on the training set. First, each image was normalized to a fixed size of 224×224 pixels. Then, the pixel values were rescaled to the range of [0, 1] by dividing by 255. During training, we used the following real-time data augmentation techniques using Keras' ImageDataGenerator: (1) Random horizontal flip: probability of 0.5. (2) Random scaling: scaling range of ±10%. (3) Gray-level normalization: processed image by image by image by normalizing the images to zero mean and unit variance.
[0033] The validation and test sets were only resized and normalized, without any data augmentation, to ensure the fairness of the evaluation.
[0034] Figure 2 This diagram displays the confusion matrix of the classification model provided by this invention performing four-class classification on a test set. The numbers on the diagonal of the matrix represent the number of correctly classified samples, while the numbers outside the diagonal represent the number of misclassified samples. Based on... Figure 2 It can be seen that the present invention performs well in the task of classifying sample data.
[0035] Table 1 shows the detailed performance of the classification model provided by this invention on four types of brain tumor datasets.
[0036] Table 1. Classification performance of the classification model provided by this invention.
[0037] As can be seen, the classification model provided by this invention has strong generalization ability in distinguishing different types of brain tumors, and can accurately differentiate normal tissue from various tumor types. Compared with other categories, the recall rate of the meningioma category is slightly lower, at 90.79%. This performance gap may be partly attributed to the relatively small sample size of this category in the training dataset.
[0038] Figure 3The ROC (Receiver Operating Characteristic) curves for four classification tasks are shown, demonstrating the superior ROC-AUC performance of the classification model provided by this invention. The AUC (Area Under the Curve) values for all categories exceed 0.99, with the AUC for no tumor, pituitary gland, and both microscopic and macroscopic averages reaching 1.00. This indicates that the classification model provided by this invention has virtually no errors in distinguishing these categories. The ROC curves are almost flush with the upper left corner of the graph, indicating the model's extremely strong classification ability. The experimental data shows that although the recall rate for the specific category of meningioma is slightly lower, overall, the classification model provided by this invention exhibits high-level performance in most categories, especially excelling in the classification of no tumor and pituitary gland. Both macroscopic and microscopic averages show an AUC value of 1.00, fully demonstrating that the model achieves an ideal balance between overall performance and specific performance in each category. The ROC curves further confirm that the classification model provided by this invention performs excellently in the brain tumor magnetic resonance image classification task. Despite a slight class imbalance in the dataset, the model maintained robust performance across most classes.
[0039] To comprehensively evaluate the performance of the classification model provided by this invention, the following experiments compare the classification model provided by this invention with existing convolutional neural network models, and the results are shown in Table 2. The convolutional neural network models used for comparison include: DenseNet121 (densely connected convolutional neural network), MSCNN (Multi-Scale CNN), VGG16 (deep convolutional neural network), ResNet18 (a lightweight version of ResNet), and EfficientNet-B3 (efficient composite scaling network).
[0040] Table 2 Comparison of experimental results for different classification models
[0041] Based on the data in Table 2, it can be seen that, compared with other models, the classification model provided by this invention has the best evaluation index and number of parameters.
[0042] In summary, Example 1 fully leverages the advantages of both multi-scale channel attention convolutional neural networks and support vector machines, providing a more accurate, efficient, and flexible solution for classifying brain tumor MRI images. Furthermore, Example 1 employs a cosine annealing combined with a warm restart strategy, which accelerates model convergence and improves generalization ability. Extensive experiments based on brain tumor MRI datasets demonstrate that the proposed scheme in Example 1 achieves competitive performance in terms of accuracy, precision, recall, and F1 score, exhibiting strong robustness and generalization ability, making it suitable for practical applications.
[0043] Example 2: Example 2 provides a classification device for brain tumor magnetic resonance images, comprising: Dataset building unit, used to acquire brain tumor magnetic resonance image datasets; The model building unit is used to build a classification model that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The model training unit is used to train the constructed classification model using the dataset to obtain the trained classification model. The image classification unit is used to input the MRI image of the brain tumor to be classified into the trained classification model and output the classification result.
[0044] The brain tumor magnetic resonance image classification device provided in Example 2 is used to perform the steps in the brain tumor magnetic resonance image classification method as described in Example 1.
[0045] Since the functions of each unit in the brain tumor magnetic resonance image classification device provided in Embodiment 2 correspond to the steps in the brain tumor magnetic resonance image classification method provided in Embodiment 1, they can be understood by referring to the description of Embodiment 1, and will not be repeated here.
[0046] In summary, the present invention has at least the following technical effects or advantages: (1) High accuracy: The classification model (i.e., MCACNN-SVM model) provided by this invention achieves a higher classification accuracy than existing methods on the Kaggle brain tumor dataset.
[0047] (2) Strong robustness: Through multi-scale convolution and channel attention mechanism, the model of the present invention can effectively improve the recognition ability of brain tumor images, reduce overfitting, and improve the stability of classification model.
[0048] (3) Low computational complexity: Although multi-scale convolution and channel attention mechanisms are introduced, the computational complexity of the model is effectively controlled, making it suitable for processing large-scale datasets.
[0049] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for classifying magnetic resonance images of brain tumors, characterized in that, Includes the following steps: Acquire a dataset of magnetic resonance images of brain tumors; A classification model is constructed that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The constructed classification model is trained using the dataset to obtain a trained classification model; The brain tumor MRI image to be classified is input into the trained classification model, and the classification result is output.
2. The method for classifying brain tumor magnetic resonance images according to claim 1, characterized in that, The images in the dataset have been preprocessed and labeled with brain tumor categories.
3. The method for classifying brain tumor magnetic resonance images according to claim 2, characterized in that, The preprocessing includes at least one of normalization and data augmentation; the normalization includes at least one of size normalization and grayscale normalization; the data augmentation includes at least one of random rotation, horizontal / vertical flipping, and random scaling.
4. The method for classifying brain tumor magnetic resonance images according to claim 1, characterized in that, The multi-scale convolutional neural network uses convolutional kernels of various sizes to extract image features in parallel.
5. The method for classifying brain tumor magnetic resonance images according to claim 4, characterized in that, The multi-scale convolutional neural network includes three parallel convolutional paths, each of which consists of a convolutional layer, a batch normalization layer, and a max pooling layer connected in sequence; the convolutional kernels of the three convolutional paths are 3×3, 5×5, and 7×7, respectively.
6. The method for classifying brain tumor magnetic resonance images according to claim 1, characterized in that, The training process of the classification model is optimized by using the hot-restart stochastic gradient descent algorithm.
7. The method for classifying brain tumor magnetic resonance images according to claim 6, characterized in that, The hot-restart stochastic gradient descent algorithm combines a stochastic gradient descent optimizer with a hot-restart mechanism; it uses cosine annealing with hot restart as a learning rate scheduler to assist the stochastic gradient descent optimizer in training the classification model.
8. The method for classifying brain tumor magnetic resonance images according to claim 1, characterized in that, Also includes: The trained classification model is evaluated for performance; wherein the performance evaluation metric includes accuracy, which is the proportion of correctly predicted samples to the total number of samples.
9. The method for classifying brain tumor magnetic resonance images according to claim 8, characterized in that, The performance evaluation metrics also include at least one of recall, precision, F1 score, macro average, and micro average; wherein, precision is the proportion of all samples that the model predicts to be positive samples, but which are actually positive samples.
10. A classification device for magnetic resonance images of brain tumors, characterized in that, include: Dataset building unit, used to acquire brain tumor magnetic resonance image datasets; The model building unit is used to build a classification model that integrates a channel attention mechanism, a multi-scale convolutional neural network, and a support vector machine; wherein the channel attention mechanism is embedded in the feature extraction module of the multi-scale convolutional neural network, and the support vector machine is connected to the feature output end of the multi-scale convolutional neural network as a classifier; The model training unit is used to train the constructed classification model using the dataset to obtain the trained classification model. The image classification unit is used to input the brain tumor MRI image to be classified into the trained classification model and output the classification result. The brain tumor magnetic resonance image classification device is used to perform the steps in the brain tumor magnetic resonance image classification method as described in any one of claims 1 to 9.