Deep Learning Based Classification of Brain Tumor Types from MRI Scans

Abstract

Patients suffering from brain tumors have significant rate of mortality. The diagnosis of brain tumors, if carried out erroneously, may result in incorrect medical intervention and hence may reduce the chance of survival of the patient. Since the risk of developing brain tumors increases with age and as there has been an increase in the aging population in the world, an urgent need is felt to develop simple and low cost analytical tools for its early diagnosis. Usually, MRI scans are used to image a patient’s brain. In recent times, machine learning and its sub-domain, deep learning has reduced the need for human judgment in the diagnosis of diseases. Deep learning models are increasingly being adopted in lieu of traditional supervised learning algorithms due to their inherent advantages owing to their ability to gather requisite details from the images automatically. One of the most difficult bio-medical imaging problems is to detect whether a patient has developed brain tumor. Despite having literatures dedicated to detect or classify various types of brain tumors through deep learning approaches, they lack high accuracy. Here, a solution for the classification of two types of brain tumors, namely Meningioma and Glioma is presented. A novel 13-layer deep convolutional neural network (CNN) architecture is used, that is built from scratch. The proposed system after performing 10-fold cross-validation gives an average validation accuracy of 100%. It is the highest attainable accuracy among existing works performed on axial MRIs, and on the same dataset.