Brain MRI Segmentation Using Autoencoders

Abstract

AbstractBrain MRI segmentation is a popular area of research that has the potential to improve the efficiency and effectiveness of brain related diagnoses. In the past, experts in this field were required to manually segment brain MRIs. This grew to be a tedious, time consuming task that was prone to human error. Through technological advancements such as improved computational power and availability of libraries to manipulate MRI formats, automated segmentation became possible. This study investigates the effectiveness of a deep learning architecture called an autoencoder in the context of automated brain MRI segmentation. Focus is centred on two types of autoencoders: convolutional autoencoders and denoising autoencoders. The models are trained on unfiltered, min, max, average and gaussian filtered MRI scans to investigate the effect of these filtering schemes on segmentation. In addition, the MRI scans are passed in either as whole images or image patches, to determine the quantity of contextual image data that is necessary for effective segmentation. Ultimately the image patches obtained the best results when exposed to the convolutional autoenocoder and gaussian filtered brain MRI scans, with a dice similarity coefficient of 64.18%. This finding demonstrates the importance of contextual information during MRI segmentation by deep learning and paves the way for the use of lightweight autoencoders with less computational overhead and the potential for parallel execution. KeywordsBrain MRISegmentationAutoencodersFiltersImage patches