Sequence-Aware and Advanced Biomarker Calculation Improves Statistical Inference in Image Processing of Parkinson’s Disease

Abstract

Improved segmentation of magnetic resonance imaging is necessary to provide quantitative and anatomical information for current and future radiological understanding of Parkinson’s disease and progression. Image segmentation is a common task, accomplished by a family of approaches, for calculation of volumetric and structural biomarkers in medical images. In Parkinson’s there is a focus on understanding subcortical grey matter, in particular as localization for deep brain stimulation surgery. In this work, we propose the following contributions on the multi-atlas segmentation framework. First, we propose an expansion of the statistical label fusion generative models to incorporate atlases of multiple labeling protocols. Second, we propose a multi-atlas segmentation framework to account for variability in imaging sequences. Third, we present a segmentation approach for efficient segmentation of the hippocampus and amygdala in the presence of a large atlas population. Fourth, we present a segmentation approach for automated segmentation of anatomical structures with highly variable anatomies, in particular the cerebellar lobules. Fifth, we present an evaluation of multi-modal segmentation of the subcortical grey matter. Sixth, we present an improved approach for assessment of segmentation accuracy and determination of the number of atlases needed for a given segmentation task. Together, this work expands and characterizes multi-atlas segmentation for use in Parkinson’s disease. These contributions improve our understanding of segmentation with varying labeling protocols and imaging sequences, our methods for evaluating segmentation results, and our methods for segmenting particular structures of interest.