Automatic Brain Tumor Classification Based on Transfer Learning Models

Abstract

It is time-consuming and error-prone to manually determine whether there is a brain tumor in an image. However, traditional automatic classification algorithms have certain limitations, which makes the automation of brain tumor classification still a challenging problem. In this article, a new method for automatic classification of brain tumors is proposed, which combines neural network models with transfer learning methods, so as to improve or solve the problem of slow iteration and long time-consuming model generation, improve accuracy, and reduce parameter. In short, the convolutional neural network model (CNN) is combined with the method of transfer learning to achieve automatic image classification on the Brain Tumor Detection 2020 dataset provided by Model Whale. More specifically, during the experiment, Tensorflow was selected as the deep learning framework. First, the transfer learning method was used, and imagenet weights were used. Then, Comparing model performance by changing the choice of the backbone network of the CNN. Select the accuracy rate as the evaluation index, compare the performance of the model, use binary_crossentropy as the loss function, and the optimizer uses adam. In this paper, three backbone networks, VGG, MobileNet and ResNet, are compared. Experimental results indicate that the automatic classification of brain tumors with the combination of CNN model and transfer learning method has better performance and the VGG model has the best performance.