Visualizing large medical image datasets using the 3D scale-invariant feature transform (SIFT)

Abstract

This thesis proposes an image classification paradigm where instead of fully automatic classification, the goal is to generate a highly-informative visual summary of class-related information for human interpretation. Rather than providing a single classification, we provide a visualisation highlighting the information relevant to group differences. In this particular research, we provide a survey of instance-based classification and visualization. A probabilistic framework is developed for classification and visualization, based on the 3D scale-invariant feature transform (SIFT) format. We propose a novel kernel density bandwidth estimator for SIFT feature densities, based on hypothesis testing, where the bandwidth minimizes the p-value of Fisher’s exact test. We also propose a method of of ranking features based on the false discovery rate (FDR). An existing implementation of the SIFT-Rank method Toews & Wells (2013) is used for feature extraction, and classification and visualization are implemented in MATLAB. We validate our approach on 3D magnetic resonance image (MRI) data of the adult human brain from the Open Access Series of Imaging Studies (OASIS) dataset Marcus et al. (2007). Experiments investigated classification and visualisation of three binary categories: age (young, old), gender (male, female), and disease (Alzheimer’s disease vs. healthy). The highest classification accuracy was 94.71% for age (old vs. young), and the method may prove useful for understanding the aging process. The method is generally applicable to arbitrary 3D medical image modalities and conditions, for example computed tomography (CT) lung scans.

Date	8 Apr 2020
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Matthew Toews (Supervisor)