Research And Implementation Of Brain Tumor Image Classification Based On Deep Learning

According to the World Health Organization,cancer has become the second largest culprit in human death,and brain cancer is one of the most lethal types of all kinds of tumors.For patients with brain tumor,it is very important to determine the type of brain tumor early for the development of special treatment plan and survival rate after treatment.Medical imaging technology is usually chosen as the first choice to distinguish the types of brain tumors.In the past,the diagnosis of brain tumors requires doctors to read tumor images,but people have limited energy.A large number of repetitive work will lead to the improvement of diagnostic error rate and aggravate the contradiction between doctors and patients.In order to overcome these problems,the computer-aided diagnosis system based on pathological image has developed rapidly.At present,deep learning has been successfully applied in the field of medical diagnosis.The accuracy of diagnosis can be improved by predicting the brain tumor image by deep learning.In order to solve the problem that traditional convolutional neural network can easily ignore the key position information of the lesion area in the image,this thesis proposes a gate control channel attention conversion unit,which uses a small number of parameters to change the relationship between channels,and improves the classification accuracy without increasing the calculation cost.A multi-path attention network is constructed by integrating the control channel attention conversion unit and the multi-path network.In order to solve the data imbalance caused by the data inconsistency of different categories in the dataset and get better classification effect,this thesis proposes a weighted loss function.Based on the combination of cross entropy loss function and mean square error loss function,weight is used to make up for the problem of too large accuracy difference among categories caused by the inconsistency of sample number of each category.Cross entropy loss function The number can ensure that the convergence speed of the model is not affected.The advantage of mean square error loss function is that the greater the difference between the predicted results and the real results,the greater the penalty.In this thesis,the mean square error loss function and cross entropy loss function are weighted to form the joint weight cross entropy loss function,and it is used as the loss function of multi-path attention network.In view of the problem of less data in brain tumor data set and the disadvantage of convolutional neural network losing space feature,the basic network is changed from convolutional neural network to capsule network.A feature attention module is proposed combining spatial attention mechanism,gate control channel attention conversion unit and channel attention mechanism.The convolution layer of capsule network is improved by using feature attention module.The high-level features extracted are encoded by the capsule layer of the capsule network,and the relationship between the extracted highlevel features and the digital capsule layer is processed by dynamic routing algorithm to enhance the expression of the capsule network.This thesis designs a classification system of brain tumor image based on deep learning,which can predict the classification according to the user provided brain tumor image.The system can help doctors make diagnosis,reduce the time of diagnosis and improve the accuracy of diagnosis.