Multi-Atlas Segmentation through Rater Performance Modeling: Theory and Applications

Abstract

The ability to generalize information from examples has been the driving force behind decades of statistical modeling and machine learning research. Building on this fundamental concept, this dissertation addresses the ability to generalize structural context, or segmentations, from medical images using labeled examples (i.e., atlases). Specifically, this research focuses on the problem of multi-atlas segmentation in which image correspondences between a set of atlases and the target-of-interest are discovered and the underlying target segmentation is estimated using statistical fusion – a supervised learning approach for resolving label conflicts. Using this general framework, several theoretical advancements to the statistical fusion model are presented, and the results of these contributions are highlighted on clinically and scientifically relevant applications. Herein, for the proposed theoretical contributions, the generative models governing statistical fusion are revisited to simultaneously and optimally account for: (1) spatially varying task difficulty, (2) spatially varying atlas performance, (3) imperfect image correspondence, and (4) hierarchically consistent performance estimation. Next, the benefits of these theoretical advancements are illustrated for: (1) detection of imaging abnormalities and anomalies, (2) segmenting the spinal cord’s internal structure through structural shape and appearance modeling, and (3) removing the need for computationally expensive deformable registration in whole-brain multi-atlas segmentation via machine learning mechanisms.