Study On MRI Image Aided Diagnosis Of Brain Tumor Based On Convolutional Neural Network

In recent years,early diagnosis and treatment of brain tumors are of great practical importance to alleviate patients’ pain and improve their survival rate.Deep learning is widely used in clinical medicine,and the identification of brain tumor types provides a valuable reference for research on related diseases.Magnetic resonance imaging(MRI)is one of the most important tools to detect brain tumors,which are irregular in shape and uneven in distribution.Most of the traditional diagnostic methods rely on experienced professional physicians to identify images with the naked eye,which is inefficient,unstable,and prone to oversight and misjudgment and aggravates the conflict between doctors and patients.The application of computer-aided means in the field of brain tumor diagnosis is one of the research hotspots.To address the above problems,this paper explores the application of computer-aided means in brain tumor diagnosis by deep learning method,taking MRI brain tumor images as the research object,and discusses the application of computer-aided means in brain tumor diagnosis from the actual brain tumor image classification needs,so as to effectively improve the accuracy of doctors’ diagnosis of the disease.Convolutional neural networks and brain tumor image recognition tasks are combined in this paper.Firstly,four convolutional neural networks,EfficientNetV2,VGG16,Dense Net169 and Res Ne Xt,are trained after preprocessing using the public open source CE-MRI dataset.Since most of the models need to increase the accuracy rate by increasing the network depth,but too many parametric numbers require high arithmetic power and can slow down the operation.To solve this problem,the EfficientNetV2 lightweight network is selected as the base model of this paper through experimental analysis based on the simultaneous consideration of model accuracy and computing speed.The EfficientNetV2 network-based brain tumor classification has the ability to automatically scale the depth,width and image resolution of the network,and this advantage effectively improves the classification accuracy.Compared with three commonly used models for brain tumor classification tasks,VGG16,Dense Net169 and Res Ne Xt,the accuracy of the validation set reaches 94.8%,and the training time is 3h21 min,compared with other models,the training time is reduced by 37%,29%,and 20%,respectively.Secondly,to solve the problem that the attention cannot be accurately focused on the lesion site of brain tumor images,which makes the classification accuracy decrease.In this paper,the coordinate attention mechanism will be introduced,which will acquire the feature information of brain tumor from both vertical and horizontal directions at the same time to accurately identify the lesion features of brain tumor and make the model classification accuracy higher.In order to further improve the classification accuracy,Hard-Swish activation function is introduced,which can not only improve the computational speed of the brain tumor classification network model,but also effectively improve the classification accuracy.Meanwhile,the improved model with Dropout and normalization layers can better suppress the occurrence of overfitting.Finally,the improved model achieves 98.4% classification accuracy in the validation set,which is3.6% better than the original network.The experimental results demonstrate that the improved model has good performance in the classification of brain tumor images,which can assist doctors in the diagnosis process with less time and effort and more accurate determination of brain tumor categories,and also provide new ideas for the auxiliary diagnosis and treatment of other brain diseases.