Interpretable Machine Learning for Brain Tumor Analysis Using MRI

Abstract

A brain tumor is a potentially fatal growth of cells in the central nervous system that can be categorized as benign or malignant. Advancements in deep learning in the recent past and the availability of high computational power have been influencing the automation of diagnosing brain tumors. DenseNet and U-Net are considered state of the art deep learning models for classification and segmentation of MRIs respectively. Despite the progress of deep learning in diagnosing using medical images, generic convolutional neural networks are still not fully adopted in clinical settings as they lack robustness and reliability. Moreover, such black-box models don’t offer a human interpretable justification as to why certain classification decisions are made, which makes them less preferable for medical diagnostics. Brain tumor segmentation and classification using deep learning techniques has been a popular research area in the last few decades but still, there are only a few models that are interpretable. In this paper, we have proposed an interpretable deep learning model which is more human understandable than existing black-box models, designed based on U-Net and DenseNet to segment and classify brain tumors using MRI. In our proposed model, we generate a heat map highlighting the contribution of each region of the input to the classification output and have validated the system using the MICCAI 2020 Brain Tumor dataset.