Convolutional Neural Network to Classify Medical Images of Rare Brain Disorders

Abstract

In recent years, the widespread dominance of convolutional neural networks (CNN) in numerous computer vision applications, particularly in medical imaging, has been compelling. However, their applications as classifiers are tedious since they need high volume (usually several hundred to several thousand) and thorough preparation of training samples to learn competently. Sometimes, it is nearly impossible to collect such a large number of unique images, especially for rare diseases (i.e., Multiple Sclerosis). Hence, we proposed a CNN that required only sixty unique and nearly unprocessed samples to learn to classify disparate samples of the same disorder with an accuracy of 85%, making it highly likely to overcome the aforementioned constraint. Although due to the paucity of patients with rare brain disorders, in this research we deployed the model to perform classifications of tumorous and hemorrhagic scans against normal ones, it could be generalized to images of other conditions, even rarer ones, since it does not require much to learn.