Keypoint masking for analyzing segmented medical image data

Abstract

This thesis investigates different methods of using segmentation masks, in the context of keypoint analysis of medical images and specifically the human brain in magnetic resonance images (MRI). Recent studies have used keypoints extracted following skull-stripping, i.e. first removing all non-brain image content. However we hypothesized that skull-stripping prior to convolution filtering (e.g. Gaussian derivative filtering used in 3D SIFT-Rank keypoint extraction) will lead to random boundary effects that will hinder brain analysis. To test this hypothesis, we compare against keypoints extracted from natural images prior to skull-stripping. Our experiment replicates a recent large-scale neuroimage family indexing experiment on data from Human Connectome Project, where classification results improve on average 2% for keypoints extracted from natural data vs. skull-stripped data. We develop a theoretical model explaining and predicting experimental results based on the properties of a n-dimensional normal distribution. Our methodology is general, and we expect our results to generalize to other non-brain data, e.g. natural image regions and other classification systems based on linear convolution, e.g. convolutional neural networks.

Date	22 Dec 2020
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Matthew Toews (Supervisor) & Rola Harmouche (Co-supervisor)