Research On Convolutional Neural Network Model For Brain Tumor Mri Classification

Brain tumor is one of the most invasive diseases today.Early detection,early diagnosis and early treatment are very important for patients.MRI is an important means in brain tumor diagnosis.With the large collection and storage of MRI images,facing the large data image database,manual interpretation and common interpretation methods cannot be processed,and the detection accuracy is not high.Therefore,it is necessary to use deep learning method to diagnose brain tumors.Convolutional neural network(CNN)is the most successful deep learning method in image classification.However,there are many CNN models.In view of the high dimensionality and complexity of actual data,it is necessary to optimize the appropriate network model and optimize the super parameters in the model.Therefore,this thesis mainly studies the MRI Classification Technology of brain tumors based on CNN from the perspective of transfer learning and Bayesian theory.The main research work is as follows:(1)Research on an improved CNN model based on VGG16.In order to solve the problems that the dense layer of VGG16 model contains a large number of parameters,requires a large amount of calculation,occupies a large amount of memory,which makes it difficult to implement in the front end,and how to scale the image to adapt to the input size of VGG16 model when using transfer learning,an improved CNN topology based on VGG16 is proposed,which includes four main modifications in order to achieve better performance.The original structure of VGG16 is significantly changed,and numerous models are built and tested on 3064 MRI brain tumor images.Setting the stride of the max pooling layers to its default value resulted in a considerable increase in overall performance.The experimental results show that the VGG16-V4 is the most effective of the four modified VGG16 models(VGG16-V1,VGG16-V2,VGG16-V3,and VGG16-V4),with 94.93% accuracy,94.18% precision,94.32%recall,and 95.02% F1-score.The use of dropout,along with a reduction in the number of filters and dense nodes,resulted in an excellent performance,which shows that the improved VGG16 topology is effective and feasible.(2)Research on CNN model hyperparameter optimization based on Bayesian optimization.The choice of hyperparameters in a CNN model has a direct impact on the model’s accuracy.Therefore,a Bayesian optimization-based model is proposed,which adopts the modified VGG16-V4 architecture and it includes the optimal choices for five hyperparameters(the activation function,the batch size,the dropout rate,the number of dense nodes,and the optimizer).On the MRI data set,this model is compared with pre-trained VGG16,VGG19,Residual Networks 50(Res Net50),Inception V3,and Densely Connected Convolutional Networks 201(Dense Net201)models.The experimental findings indicate that training CNN from scratch outperforms Transfer Learning for the classification of brain tumors.The diagnostic accuracy of the optimized VGG16-V4 can reach 98.70%,while the basic VGG16,VGG19,Res Net50,Inception V3,and Dense Net201 achieved 97.08%,96.43%,89.29%,92.86%,and 94.81% diagnostic accuracy,respectively.Moreover,the performance of the proposed model on CE-MRI data set is superior to that of the popular methods,which shows that the hyperparametric optimization based on Bayesian optimization is feasible and effective.