All ETDs from UAB

Advisory Committee Chair

Murat M Tanik

Advisory Committee Members

Leon Jololian

Kent Keyser

Buren E Wells

Document Type

Dissertation

Date of Award

2020

Degree Name by School

Doctor of Philosophy (PhD) School of Engineering

Abstract

The Parkinson's Progression Markers Initiative (PPMI) dataset (which includes several types of imaging datasets) was created to generate biomarkers for diagnosing the presence and severity of Parkinson’s Disease (PD). Unlike some conditions, PD is not intuitively diagnosed directly from brain imaging. As such, it is vital to understand as well as classify PD. Deep learning techniques are becoming increasingly prevalent in medical imaging workflows. In some cases, deep learning outperforms human diagnostic ability. However, the black-box nature of many machine learning models can be troubling, particularly in a medical context since the results could imply alternate treatments and a human caregiver must ultimately take responsibility for diagnosis and treatment. Black-box algorithms provide only results, with no information regarding the process leading to that result. The work presented in this dissertation advances machine learning techniques through developing an augmented framework for comparing machine learning predictions with statistical methods of classification currently used in medical imaging. This provides a basis for determining if deep learning architectures, pre-processing steps, or dataset augmentations have desired effects on predictions. Developing a framework for a) validating machine learning models against other techniques and b) describing what the machine learning model prediction is based off of can be beneficial for adding transparency to a traditionally opaque process. As a second contribution of the presented dissertation, a novel, efficient, model-agnostic, and visualization method of black box models was developed. This novel visualization method is a novel approach to occlusion sensitivity which adds the notion of hierarchy and reduces the iterations required by three orders of magnitude. For PD, using deep learning workflows which treat a model as a black box have no potential to develop biomarker information through training and classification alone. Using the framework presented in this dissertation, PPMI datasets were used to train a PD classification model. The framework and visualization method were then applied to the resultant machine learning model to augment the classification provided with population level information from visualizations and voxel-wise statistics to provide a basis for developing information towards biomarker discovery and model validation, beyond simply developing a classifier.

Included in

Engineering Commons

Share

COinS