Classification of Brain Tumor MRI Scans using Transfer Learning with a Comparative Analysis on Pre-Trained Networks

Abstract

Classification of brain MRI scans accurately to determine whether they are tumorous or not is a challenging task. Deep learning models are popularly used for the classification of data and requires lot of computational time and memory, however flexibility is highest when trained from scratch for a particular application. Where there is a constraint in availability of datasets, memory, computational time and efforts, transfer learning plays a major role to train models efficiently. In this work, transfer learning is used to classify the brain MRI scans to determine whether they are tumorous or not. Transfer learning helps to train the pre-trained models on a smaller dataset by fine-tuning the last learnable layers. Pre-trained models like GoogleNet, Alexnet, SqueezeNet, VGG16 and VGG19 are first trained and tested on a small ‘Brain Tumor’ dataset. Training is performed by fine-tuning the learnable layer of each network and by setting the training options. Performance is evaluated by using confusion matrices, SqueezeNet achieved highest accuracy of 92.08%. Further Alexnet and SqueezeNet are trained on the dataset consisting of 3064 T1-weighted images. Thus, modifying and fine-tuning the last layers, transfer learning helps to train a network which was pre-trained on a different dataset to be further trained on the desired dataset for making predictions. Comparative analysis is presented at the end of the study.