Finding the Most Effective Data Augmentation Techniques on Brain MRI Data Using Deep Networks

Abstract

In 2020, over 250,000 people died from brain and Central Nervous System (CNS) tumors. Brain tumors account for around 85% or more of these. As of 2021, over 16 million people in the United States have been diagnosed with some type of cognitive impairment. The goal of this paper is to find the most effective set of data augmentations to correctly classify cognitive diseases with deep networks, using structural Magnetic Resonance Imaging (MRI) data. This paper demonstrates a Greedy optimization technique to find the most effective sequence of data augmentations out of blurring, distortion, position, and red noise (an overlay augmentation displaying random clouds of noise on the images). We sought to classify three tumors: glioma tumors, pituitary tumors, and meningioma tumors, as well as detect if there was no tumor at all. We also classified three stages of Alzheimer’s disease: not demented, very mildly demented, and mildly demented, to further demonstrate the effectiveness of the data augmentation sequence.