Classification of brain tumors from MRI images using deep transfer learning

Abstract

Each year, more than 100,000 people in the United States are diagnosed with a brain tumor. An early and accurate diagnosis is crucial in getting patients the necessary treatment and increasing survival rates. In recent years, machine learning algorithms have become increasingly popular in the medical field due to their ability to recognize complex patterns and reduce human errors. However, accurate diagnosis using deep learning algorithms requires a large amount of training data, which is not always available. Additionally, training a model from scratch can take a long time and requires vast amounts of computational power. As a solution, this study aims to utilize transfer learning, which uses the knowledge gained by the model on one dataset to aid in classifying the second dataset. In this study, a dataset of MRI images consisting of four classes of brain tumors (no tumor, pituitary tumor, meningioma, and glioma) were used. The performance of seven pre-trained models (ResNet18, ResNet50, VGG16, DenseNet, GoogLeNet, ShuffleNet, and MobileNet) were evaluated in order to see which would achieve the highest classification accuracy. Additionally, this study examined two different methods for the implementation of transfer learning. In the first method, the convolutional base of the pre-trained model was frozen and in the second method, the convolutional base was trained. The best performing models proved to be ResNet18 and ShuffleNet with the base trained, achieving an accuracy of 97.86%. The results also showed that the models with the convolutional base trained outperformed those with the convolutional base frozen.