MAGNETIC RESONANCE IMAGING BIOMARKERS FOR PARKINSON’S DISEASE: A MACHINE LEARNING APPROACH
Abstract
Parkinson’s disease (PD) is a progressive neurodegenerative disorder that primarily affects the elderly population, characterized by motor symptoms and non-motor symptoms. Unfortunately, there are currently no curative treatments for PD, largely due to the absence of reliable diagnostic biomarkers, particularly in the early stages of the disease. Early detection plays a crucial role in timely intervention and enrollment in clinical trials. This thesis project focuses on the development of robust biomarkers for PD using brain structural morphometry, with a specific emphasis on the striatum and ventral tegmental area/substantia nigra compacta (VTA/SNc), which are prominently affected in PD. The first chapter provides an in-depth back- ground on PD, existing biomarkers, and their limitations. Furthermore, it presents the methodology employed for the proposed analysis.
A novel automated parcellation pipeline called is introduced in the second chapter, which subdivides subcortical structures into distinct regions based on their connectivity to the cortex. This pipeline enables the extraction of various features related to volume, surface characteristics, and connectivity, enabling the identification of abnormalities in patient populations. This chapter evaluates the reliability and diagnostic capability of these features.
In the third chapter, machine learning models are trained to detect early-stage PD using features extracted from subregions of the striatum and VTA/SNc derived from multicentered MRI data. These models exhibit high accuracy and demonstrate that specific features, such as surface-based and bundle-based fractional anisotropy and mean diffusivity, play a crucial role in the classification of PD. The parcellation of subregions leads to an improvement in PD detection compared to unparcellated measures. The developed diagnostic test integrates shape- based measures and machine learning techniques, and ongoing validation is being conducted through a multicentered trial.
The final chapter introduces a novel approach to predict disease prognosis in PD by developing subject-specific models utilizing baseline MRI and clinical data. The proposed model, leveraging MRI data, outperforms ones that exists, demonstrating robustness in predicting severity scores and addressing assessment challenges.
This thesis advances PD research by developing reliable biomarkers, enhancing diagnostic accuracy, and predicting disease prognosis.