Brain tumor classification method based on improved OfficientNet network model

Abstract

The invention belongs to the technical field of image analysis, and particularly relates to a brain tumor classification method based on an improved OfficientNet network model. According to the method, an OfficientNet network model is improved, and a local attention enhancement module for dynamically adjusting channel importance of different regions in a brain tumor medical image by using a local region importance weighting mechanism is added at an input position of the OfficientNet network model; a to-be-classified brain tumor medical image is sequentially processed by the ECA self-attention layer, the first feature extraction layer, the soft pooling layer, the second feature extraction layer, the first activation layer and the feature size adjustment layer to obtain a first feature having the same size as the to-be-classified brain tumor medical image; and the feature optimization layer performs feature optimization based on the first feature and the brain tumor medical image to be classified, and inputs an optimization result into the OfficientNet network model to obtain a brain tumor classification result, so that the classification precision of the model is enhanced while the calculation complexity is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image analysis, and particularly relates to a brain tumor classification method based on an improved EfficientNet network model. Background Art

[0002] Brain tumors are a very serious type of cancer that pose a great threat to humans and are difficult to diagnose. With the rapid development of artificial intelligence technology, the application of deep learning-based medical image analysis technology in brain tumor detection and classification has received extensive attention. It supports clinical diagnosis and treatment planning by automatically analyzing brain tumor medical images (such as magnetic resonance imaging MRI or computed tomography images CT) to identify and classify tumor types. Among them, convolutional neural network models gradually learn the key features of tumors in images through hierarchical feature extraction and have achieved good results in brain tumor classification tasks, and have been widely used in brain tumor classification tasks.

[0003] For example, in 2016, a brain tumor classification method based on the Res Net network model was proposed, achieving a relatively high classification accuracy on the standard MRI dataset. However, when using the Res Net network model or the Dense Net network model for brain tumor classification, the model has a large number of parameters and a high computational complexity, usually requiring high-performance hardware support and being difficult to adapt to resource-constrained clinical scenarios. Transfer learning accelerates the model training process and reduces operating costs by leveraging models pre-trained on large-scale datasets (such as ImageNet) and transferring their feature extraction capabilities to practical applications. Therefore, existing research has proposed transferring models pre-trained on large-scale datasets to medical image analysis for brain tumor classification. However, transfer learning has limited adaptability to specific tasks, cannot fully utilize the characteristics of medical images, has low classification accuracy, and has a computational efficiency bottleneck when processing high-resolution MRI images. With the development of deep learning in the field of computer vision, in recent years, image classification models such as Vision Transformer and Swin Transformer have emerged. These models have high computational complexity, strong dependence on large-scale training data, and a large number of model parameters, and their accuracy performance in some brain tumor classification tasks is not as good as that of traditional convolutional neural networks. Therefore, existing brain tumor classification models cannot achieve both high classification accuracy and low computational complexity simultaneously. Summary of the Invention

[0004] The purpose of the present invention is to provide a brain tumor classification method based on an improved EfficientNet network model to solve the problem that existing brain tumor classification models cannot achieve both high classification accuracy and low computational complexity simultaneously.

[0005] The present invention provides a brain tumor classification method based on an improved EfficientNet network model to solve the above technical problems, including: inputting a brain tumor medical image to be classified into the improved EfficientNet network model to obtain a brain tumor classification result; wherein, the improved EfficientNet network model includes a local enhanced attention module and an EfficientNet network model; the local enhanced attention module uses a local region importance weighting mechanism to dynamically adjust the channel importance of different regions in the brain tumor medical image, including an ECA self-attention layer, a first feature extraction layer, a second feature extraction layer, a soft pooling layer, a first activation layer, a feature size adjustment layer, and a feature optimization layer. The brain tumor medical image to be classified is sequentially processed by the ECA self-attention layer, the first feature extraction layer, the soft pooling layer, the second feature extraction layer, the first activation layer, and the feature size adjustment layer to obtain a first feature with the same size as the brain tumor medical image to be classified. The feature optimization layer performs feature optimization based on the first feature and the brain tumor medical image to be classified, and inputs the optimization result into the EfficientNet network model to obtain a brain tumor classification result.

[0006] Further, the feature optimization layer includes a gated network layer and a second activation layer. The input of the second activation layer is the data of the first channel of the brain tumor medical image to be classified. The gated network layer processes the output of the second activation layer, the first feature, and the brain tumor medical image to be classified to obtain an optimization result.

[0007] Further, the formula used by the gated network layer for data processing is:

[0008] A(X) = σ(X [0] ) × ψ(σ(L(X))) × X

[0009] wherein, A(X) is the optimization result obtained after the gated network layer processes data, X [0] is the data of the first channel of the brain tumor medical image, X is the brain tumor medical image, σ(·) represents activation, and ψ(·) represents bilinear interpolation.

[0010] Further, the feature size adjustment layer uses bilinear interpolation to adjust the input feature size to the same first feature as the brain tumor medical image to be classified.

[0011] Further, the second feature extraction layer includes a strided convolution and a normal convolution. The output of the soft pooling layer is processed by the strided convolution and then input into the normal convolution, and the output of the normal convolution is connected to the input of the first activation layer; the strided convolution refers to a convolution with a stride greater than 1, and the normal convolution refers to a convolution with a stride of 1.

[0012] Further, the EfficientNet network model is an EfficientNetV2-S network model.

[0013] Further, the activation functions used in the first activation layer and the second activation layer are sigmoid activation functions.

[0014] Further, the dataset used in training the improved EfficientNet network model is preprocessed data, and the preprocessing includes size adjustment, pixel normalization, and affine transformation.

[0015] Further, when training the improved EfficientNet network model, the Adam optimizer is used to optimize the hyperparameters in the improved EfficientNet network model.

[0016] The beneficial effects of the above technical solutions are as follows: The present invention is an improved invention. A local enhancement attention module is added at the very beginning of the EfficientNet network model. In the local enhancement attention module, an ECA self-attention layer is set to adaptively weight the brain tumor medical images to be classified. After feature extraction of the adaptively weighted images, they are input into the soft pooling layer for processing. The local importance of different regions of the image is calculated using the regional Softmax of soft pooling, and then feature extraction is performed on it and the extracted features are restored to the size of the brain tumor medical images to be classified. The feature optimization layer optimizes the features based on the above processing results and the brain tumor medical image data to be classified to reduce the error of local importance, reduce the artifacts caused by feature extraction and size adjustment, improve the accuracy of feature selection, and the processing process is simple, avoiding complex calculations on high-dimensional feature maps. Then, the features processed by the local enhancement attention module are input into the EfficientNet network model for brain tumor classification. Since the EfficientNet network model itself is a lightweight model, and by introducing a local enhancement attention module on it, useful information is adaptively enhanced and useless information is suppressed according to the relative importance of the input information, and the processing process is relatively simple compared with the traditional self-attention mechanism. Therefore, the improved EfficientNet network model of the present invention not only reduces the computational complexity but also enhances the classification accuracy, enabling the EfficientNet network model to obtain a higher accuracy while running efficiently. Description of the Drawings

[0017] Figure 1 It is a schematic diagram of the improved EfficientNet network model according to the method embodiment of the present invention;

[0018] Figure 2 It is a schematic diagram of the local enhancement attention module according to the method embodiment of the present invention;

[0019] Figure 3 It is a schematic diagram of the ECA self-attention layer in the method embodiment of the present invention;

[0020] Figure 4 It is a schematic diagram of the soft pooling layer in the method embodiment of the present invention;

[0021] Figure 5 It is a construction flow chart of the improved EfficientNet network model in the method embodiment of the present invention. Detailed implementation manners

[0022] In order to make the objectives, technical solutions and advantages of the present invention clearer, the following further explains the detailed implementation manners of the present invention with reference to the accompanying drawings.

[0023] The present invention adds a LICA (Local Importance-based Channel Attention) attention module at the very beginning of the EfficientNet network model. In the LICA attention module, based on a simple network structure, useful features are adaptively enhanced and useless features are weakened according to the relative importance of the input information, which improves the accuracy of brain tumor classification while increasing the model calculation complexity. It is very suitable for resource-constrained clinical scenarios, and thus can be directly applied to the classification and recognition of brain tumor medical images, providing good diagnostic assistance for clinicians.

[0024] Method embodiment

[0025] A brain tumor classification method based on an improved EfficientNet network model of the present invention is described in detail below.

[0026] 1. Prepare a dataset for training the improved EfficientNet network model.

[0027] The dataset for training the improved EfficientNet network model is a brain tumor medical image dataset, including medical images of three specific types of tumors, namely glioma, pituitary tumor, and meningioma.

[0028] Preferably, in order to evaluate the performance of the trained improved EfficientNet network model, the present invention divides the dataset into a training set and a test set according to a ratio. For example, the dataset is divided into a training set and a test set according to a ratio of 8:2. 80% of the data is used to train the improved EfficientNet network model, and the remaining 20% of the data is reserved as an independent test set to verify the generalization ability of the improved EfficientNet network model.

[0029] Preprocess the dataset. Among them, the preprocessing of the dataset includes:

[0030] 1) Size adjustment: Adjust the sizes of the images in the dataset. For example, uniformly adjust the size of the training images to 300×300, and uniformly adjust the size of the images in the test dataset to 384×384.

[0031] 2) Pixel normalization: Normalize the pixel values of each channel in the dataset images, with a mean of [0.5, 0.5, 0.5] and a standard deviation of [0.5, 0.5, 0.5].

[0032] In addition, the present invention also performs various affine transformations on the images of the training dataset to expand the diversity of the data. This includes operations such as randomly cropping, randomly rotating, and horizontally flipping the images of the training dataset to simulate the tumor characteristics under different perspectives and imaging conditions, so that the model can learn richer patterns, increase the robustness and generalization ability of the model, and at the same time effectively reduce the risk of overfitting. And to ensure a fair comparison during the testing process and make the evaluation of the model performance more reliable, no affine transformation is performed on the test dataset. Instead, the test images are only adjusted to the same size by means of central cropping, so that all test images have the same input size, while trying to retain the central information of the original images without introducing additional artificial change factors.

[0033] 2. Construct an improved EfficientNet network model.

[0034] The improved EfficientNet network model is improved by adding a LICA module to the backbone of the EfficientNet network model. As Figure 1 shown, the improved EfficientNet network model includes a local enhanced attention module and an EfficientNet network model. The local enhanced attention module is set at the very beginning of the EfficientNet network model, that is, the output of the local enhanced attention module is connected to the input of the EfficientNet network model.

[0035] 1) EfficientNet network model.

[0036] The EfficientNet network model is a lightweight convolutional neural network model that uniformly scales depth, width, and resolution, enabling efficient image processing while reducing computational and parameter quantities while maintaining accuracy. The EfficientNet model is mainly composed of stacked Fused-MBConv (fused mobile convolution) and MBConv (mobile convolution) modules. The EfficientNet network model can adopt the EfficientNetV2 version or the latest version of the EfficientNet network model according to technological development. The EfficientNetV2 version has multiple variants, including S (Small), M (Medium), and L (Large), etc. Since the EfficientNetV2-S version model is smaller and more suitable for clinical applications, as a preferred implementation, the EfficientNet network model used in the present invention is the EfficientNetV2-S network model.

[0037] 2) Local enhanced attention module.

[0038] By introducing an importance weighting mechanism for local regions, the local enhanced attention module can dynamically adjust the channel importance of different regions in the image, enabling the network model to adaptively enhance useful features and weaken useless features, thereby more precisely capturing the key features in brain tumor medical images, improving the accuracy of feature selection, and enhancing the expressive ability of the EfficientNet network model, enabling the improved EfficientNet network model to adapt to more complex image classification.

[0039] Specifically, as Figure 2 shown, the local enhanced attention module includes an ECA self-attention layer, a first feature extraction layer, a second feature extraction layer, a soft pooling layer, a first activation layer, a feature size adjustment layer, and a feature optimization layer. The brain tumor medical image to be classified is sequentially processed by the ECA self-attention layer, the first feature extraction layer, the soft pooling layer, the second feature extraction layer, the first activation layer, and the feature size adjustment layer to obtain a first feature with the same size as the brain tumor medical image to be classified. The feature optimization layer optimizes the features based on the first feature and the brain tumor medical image to be classified, and inputs the optimization result into the EfficientNet network model to obtain the brain tumor classification result.

[0040] Among them, the ECA self-attention layer is used to adaptively weight the input brain tumor medical image using the attention mechanism to improve the information extraction ability, as Figure 3As shown in the figure, it includes: global average pooling, which performs average pooling on the feature maps of each channel to obtain a global feature vector (a single value) to represent the context information of the entire image; one-dimensional convolution to learn the associations between channels; and a sigmoid activation function to activate the convolution output to obtain the weight coefficients for each channel, which are used to weight the feature maps of each channel.

[0041] The first feature extraction layer is a normal convolution (Conv1), where normal convolution refers to convolution with a stride of 1. Through 1×1 convolution, the number of channels of the feature map is reduced, reducing the computational amount.

[0042] The SoftPool layer is as Figure 4 shown. SoftPool is a pooling with an exponential weighting mechanism that uses regional Softmax to non-linearly calculate the local importance of pixel x in the surrounding region R of the feature map. In the present invention, the stride of the pooling is set to 3, and it can also be set to other values according to actual needs.

[0043] The second feature extraction layer includes a strided convolution and a normal convolution. The output of the SoftPool layer is processed by the strided convolution to extract local statistical information and generate a preliminary importance map L(X). In this embodiment, the strided convolution is a 3×3 convolution, and the corresponding preliminary importance map L(X) is:

[0044] L(X) = Conv3×3(SoftPool(X′))

[0045] where X′ is the feature map input to the SoftPool layer.

[0046] The input after being processed by the strided convolution is input to a normal convolution, and the output of the normal convolution is connected to the input of the first activation layer to adjust the number of channels and extract features while keeping the number of channels unchanged. Strided convolution refers to convolution with a stride greater than 1. For example, in the present invention, according to the size of the image output after SoftPool, in order to improve the processing efficiency, the stride of the strided convolution (Conv3, stride = 2) is set to 2. Then, a normal convolution (Conv3) is used again for feature extraction to ensure the accuracy of feature extraction.

[0047] The first activation layer can use a sigmoid activation function or a Relu activation function, etc. to activate the convolution.

[0048] The feature size adjustment layer is used to adjust the input feature size to the same size as the first feature of the brain tumor medical image to be classified. Feature size adjustment methods such as deconvolution and bilinear interpolation can be used. Preferably, in the present invention, bilinear interpolation is used to adjust the feature size output by the activation layer in two directions to output the first feature with the same size as the brain tumor medical image to be classified.

[0049] The feature optimization layer uses a gating mechanism for feature optimization, which is used to recalibrate local importance and reduce artifacts caused by strided convolution and bilinear interpolation. And different from the existing gating units that rely on additional networks, the present invention directly uses the first channel X[0] of the medical brain tumor image to be classified as the gating signal, thereby simplifying the processing process. Specifically, the feature optimization layer includes a gating network layer and a second activation layer. The input of the second activation layer is the data of the first channel of the medical brain tumor image to be classified. The gating network layer processes the output of the second activation layer, the first feature, and the medical brain tumor image to be classified to obtain an optimized result. Preferably, the activation function used in the second activation layer is the sigmoid activation function. As other embodiments, the Relu activation function can also be used.

[0050] The formula used by the gating network layer for data processing is:

[0051] A(X) = σ(X [0] ) × ψ(σ(L(X))) × X

[0052] where A(X) is the optimized result obtained after the gating network layer processes the data, X [0] is the data of the first channel of the medical brain tumor image, X is the medical brain tumor image, σ(·) represents activation, and ψ(·) represents bilinear interpolation.

[0053] 3. Train the improved EfficientNet network model.

[0054] The constructed improved EfficientNet network model can be trained using the original training dataset. Preferably, the improved EfficientNet network model is trained using the preprocessed training dataset. During the training process, the performance of the model is monitored in real time, the neural network weights with the highest accuracy are recorded, and they are saved for subsequent use.

[0055] Among them, the hyperparameters during training include: the batch size is set to 16, the number of training epochs (Epoch) is set to 100, the early stopping rounds (early stopping) is set to 5, the initial learning rate is set to 0.0001, and the weight decay coefficient is set to 5×10 -4 . The hyperparameters in the network model are optimized using an optimization algorithm. The optimization algorithm can use the Adam optimizer, or other existing optimization algorithms, such as the L-BFGS algorithm, the AdaDelta algorithm, etc.

[0056] The weights of the neural network are initialized using the Kaiming initialization method to ensure the consistency of the variance of each layer's output, thereby accelerating the training process of the model. The initialization method is implemented using the following formula:

[0057]

[0058] Among them, W is the weight matrix; n in is the number of neurons in the previous layer (i.e., the number of input nodes); represents sampling weights from a normal distribution with a mean of 0 and a variance of

[0059] The cosine annealing learning rate adjustment strategy is adopted during the training process. The learning rate is gradually decreased during the training process, and finally the learning rate is close to the minimum value to help the model converge stably in the later stage of training.

[0060] 4. Use the test data set to evaluate the trained improved EfficientNet network model, evaluate its classification effect on unseen brain tumor medical images, record and calculate various evaluation indicators, including accuracy, precision, recall, and F1 score, to comprehensively evaluate the performance of the model. The construction process of the improved EfficientNet network model is as Figure 5 shown.

[0061] 5. Input the brain tumor medical images into the improved EfficientNet network model to obtain the brain tumor classification results.

[0062] Next, the improved EfficientNet network model (EfficientNetV2-S+LICA) proposed in the present invention is compared and analyzed with the EfficientNetV2-S network model, GhostNetV3 model, ResNext50, and Swin Transformer model in terms of the performance of brain tumor classification. The obtained data are shown in Tables 1 and 2. It can be seen that the network model designed in the present invention has achieved a significant improvement in classification accuracy compared with the unimproved EfficientNetV2-S network. Compared with traditional algorithms, the accuracy, precision, recall, and F1 score have also been significantly improved.

[0063] Table 1

[0064] Method Accuracy Precision Recall F1 Score EfficientNetV2-S 96.9％ 96.23％ 97.23％ 96.68％ EfficientNetV2-S+LICA 99.67％ 99.53％ 99.77％ 99.65％

[0065] Table 2

[0066] Method Accuracy Precision Recall F1 Score GhostNetV3 95.8％ 95.08％ 95.7％ 95.37％ ResNext50 97.10％ 96.68％ 96.92％ 96.79％ Swin Transformer 88.4％ 89.22％ 85.41％ 86.86％ VGG16 93.80％ 94.05％ 93.79％ 93.79％ EfficientNetV2-S+LICA 99.67％ 99.53％ 99.77％ 99.65％

[0067] ​By introducing the LICA attention module with a local region importance weighting mechanism into the EfficientNet network model, the present invention can dynamically adjust the channel importance of different regions of an image, thereby accurately capturing key features in brain tumor medical images. While improving the classification accuracy of brain tumor medical images, the LICA attention module significantly enhances the expressive power of the model and can adapt to more complex image content. Moreover, the LICA attention module of the present invention calculates the local importance of the input features and recalibrates the attention map in combination with the channel gating mechanism, thereby realizing second-order information interaction. Compared with the traditional self-attention mechanism, it not only improves the classification accuracy but also significantly reduces the computational complexity. By calculating the local importance on the downsampled feature map and using simple operations to optimize the feature map, while enhancing the expressive power of the model, it effectively reduces the computational latency. Therefore, the present invention significantly reduces the consumption of computing resources while ensuring high accuracy, has high application value, and is particularly suitable for brain tumor medical image analysis scenarios with high requirements for real-time performance and accuracy.

Claims

1. A brain tumor classification method based on an improved EfficientNet network model, characterized in that, Including: Input the medical image of the brain tumor to be classified into the improved EfficientNet network model to obtain the brain tumor classification result; among them, the improved EfficientNet network model includes a local enhanced attention module and an EfficientNet network model; the local enhanced attention module uses a local region importance weighting mechanism to dynamically adjust the channel importance of different regions in the medical image of the brain tumor, including an ECA self-attention layer, a first feature extraction layer, a second feature extraction layer, a soft pooling layer, a first activation layer, a feature size adjustment layer, and a feature optimization layer. The medical image of the brain tumor to be classified is processed by the ECA self-attention layer, the first feature extraction layer, the soft pooling layer, the second feature extraction layer, the first activation layer, and the feature size adjustment layer in sequence to obtain a first feature with the same size as the medical image of the brain tumor to be classified. The feature optimization layer performs feature optimization based on the first feature and the medical image of the brain tumor to be classified, and inputs the optimization result into the EfficientNet network model to obtain the brain tumor classification result.

2. The brain tumor classification method based on the improved EfficientNet network model according to claim 1, wherein, The feature optimization layer includes a gating network layer and a second activation layer. The input of the second activation layer is the data of the first channel of the medical image of the brain tumor to be classified. The gating network layer processes the output of the second activation layer, the first feature, and the medical image of the brain tumor to be classified to obtain the optimization result.

3. The brain tumor classification method based on the improved EfficientNet network model according to claim 2, characterized in that, The formula used by the gating network layer for data processing is: A(X) = σ(X [0] ) × ψ(σ(L(X))) × X Among them, A(X) is the optimized result obtained after the gated network layer processes the data, where X [0] is the data of the first channel of the brain tumor medical image, X is the brain tumor medical image, σ(·) represents activation, and ψ(·) represents bilinear interpolation.

4. The brain tumor classification method based on the improved EfficientNet network model according to claim 1 or 2, characterized in that, The feature size adjustment layer uses bilinear interpolation to adjust the input feature size to the first feature with the same size as the medical image of the brain tumor to be classified.

5. The brain tumor classification method based on the improved EfficientNet network model according to claim 1 or 2, characterized in that, The second feature extraction layer includes a strided convolution and a normal convolution. The output of the soft pooling layer is processed by the strided convolution and then input into the normal convolution. The output of the normal convolution is connected to the input of the first activation layer; the strided convolution refers to a convolution with a stride greater than 1, and the normal convolution refers to a convolution with a stride of 1.

6. The brain tumor classification method based on the improved EfficientNet network model according to claim 1 or 2, characterized in that, The EfficientNet network model is an EfficientNetV2-S network model.

7. The brain tumor classification method based on the improved EfficientNet network model according to claim 2, characterized in that, The activation functions used by the first activation layer and the second activation layer are sigmoid activation functions.

8. The brain tumor classification method based on the improved EfficientNet network model according to claim 1, characterized in that The dataset used by the improved EfficientNet network model during training is preprocessed data, and the preprocessing includes size adjustment, pixel normalization, and affine transformation.

9. The brain tumor classification method based on the improved EfficientNet network model according to claim 1, characterized in that The improved EfficientNet network model uses an Adam optimizer to optimize the hyperparameters in the improved EfficientNet network model during training.

Images