Segmentation and Severity Classification of Dementia in Magnetic Resonance Imaging using Deep Learning Networks

Abstract

Magnetic resonance imaging is the accepted standard for analyzing any deformation in brain. There are many biomarkers which can be considered for analyzing the effect of Alzheimer’s disease in brain. One such biomarker is the ventricle which expands during the progression of Alzheimer’s disease. Ventricle segmentation plays a vital role in the diagnosis. Automated segmentation approaches are preferred since manual segmentation takes a longer time. In this work, the magnetic resonance images are skull stripped using a combination of Fuzzy C-means clustering and the Chan-Vese contouring technique. segmentation of ventricle is performed by deep learning architectures, U-Net and SegUnet on 1164 transverse MR images acquired from ADNI (Alzheimer’s DiseaseNeuroimaging Initiative) database which is an open-source database for carrying researches on Dementia. The features are extracted from the segmented images using ResNet-101 and they are classified using a classifier merger approach which consists of 3 classifiers. The final class label is obtained by majority voting on the individual classifier predictions. The results were compared and analyzed.