Research On Classification Of Brain Tumor Mri Image Based On Deep Learning

Brain tumors are one of the top ten malignant tumors in morbidity and mortality today,and timely detection and accurate determination of tumor type can buy valuable treatment time for patients.Usually,the diagnosis of brain tumors is mainly marked and diagnosed by doctors based on the patient’s brain magnetic resonance imaging(MRI).However,as the workload increases,this also increases the physical and mental burden of doctors,which may affect the efficiency of diagnosis.In recent years,with the continuous development of deep learning,the use of computer vision technology to assist doctors in interpreting medical images can reduce the burden of doctors and improve diagnostic efficiency.Therefore,it is of great application value to carry out the research of brain tumor MRI image classification based on deep learning.In this paper,the classification of brain tumor MRI images was carried out and the following work was completed:In terms of data preprocessing,in order to enhance the generalization and robustness of the model,the training set in the dataset is further expanded.Aiming at the problems of unclear imaging and slow fitting speed of traditional DCGAN(Deep Convolution Generative Adversarial Networks,DCGAN),this paper proposes an image generation model based on improved DCGAN Improved-DCGAN to augment the original data set.In order to improve the clarity of the generated image,the model increases the number of convolutional layers of the generator on the basis of the original DCGAN,and introduces the method of feature fusion in the discriminator.After comparative experiments,it is proved that compared with the original DCGAN,the proposed Improved-DCGAN has a higher fitting degree of loss curve,shorter fitting time and clearer images.In terms of image classification model,in order to accurately classify brain tumor MRI image types,this paper proposes a SEAC-Res Net model,which adds Squeeze and Excitation(SE)modules with channel attention mechanism to each residual block on the basis of the Res Net backbone model,and adds Asymmetric Convolutions(AC)before the fully connected layer module,which is then classified using the Softmax classifier.By constructing the association between different feature channels,the SE module improves the importance of channels containing feature information,strengthens the directivity of network extraction of tumor regional features,and improves the attention of tumor tissues in the whole image.The AC module extracts different branch features through convolution of different shapes,increases the diversity of features,obtains more adequate features,and enhances the overall model.In order to verify the validity of the proposed SEAC-Res Net model.Ablation experiments,comparison experiments,robustness experiments and generalization experiments were carried out on the Kaggle dataset and CE-MRI dataset based on a variety of evaluation indexes(precision,recall,F1 score,specificity,accuracy,Kappa coefficient).The results show that the SEAC-Res Net model has better classification effect than the other seven brain tumor classification models(CNN-LSTM、Deep-CNN、Multiscale-CNN、HQC-CNN,Patch-Res Net、Block-Wise VGG19、RNGAP、CNN-SVM),and has strong robustness and generalization,indicating that the model can help doctors reduce their work burden as an auxiliary tool for medical diagnosis in actual medical scenarios,it is a feasible and reliable classification model for brain tumor MRI images.