A Highly Accurate Attention-Based Convolutional Neural Network for Classification of Brain Tumors

Abstract

Brain tumors have always been one of the common tumors threatening human life safety. At present, there are still relatively few computer-aided diagnostic systems in China specifically for detection of specific conditions of brain tumor, as well as related studies. This study collected a certain number of publicly available datasets of brain magnetic resonance imaging (MRI) images and data preprocessing such as normalization was conducted on it. According to the characteristics of medical image complexity of brain MRI, this study proposed an approach of incorporating attention mechanism with Convolutional Neural Network (CNN) to reduce the influence caused by irrelevant background information features in images. The experiment results based on the proposed method were compared with self-defined classic models such as VGGNet and MobileNet. Through testing on the dataset, the results show that the CNN model's accuracy after adding an attention mechanism improves significantly compared to the other three models, demonstrating that the attention mechanism in the model can reduce the impact of context irrelevant information to the classification outcome to some extent and performed well on the brain tumor recognition classification task. Finally, this paper deploys the trained analysis model on the web page, the interface is simple and friendly, and convenient for medical staff to operate.