Segmentation of medical images under topological constraints

Abstract

Major advances in the field of medical imaging over the past two decades have provided physicians with powerful, non-invasive techniques to probe the structure, function, and pathology of the human body. This increasingly vast and detailed amount of information constitutes a great challenge for the medical imaging community, and requires significant innovations in all aspect of image processing. To achieve accurate and topologically-correct delineations of anatomical structures from medical images is a critical step for many clinical and research applications. In this thesis, we extend the theoretical tools applicable to the segmentation of images under topological control, apply these new concepts to broaden the class of segmentation methodologies, and develop generally applicable and well-founded algorithms to achieve accurate segmentations of medical images under topological constraints. First, we introduce a digital concept that offers more flexibility in controlling the topology of digital segmentations. Second, we design a level set framework that offers a subtle control over the topology of the level set components. Our method constitutes a trade-off between traditional level sets and topology-preserving level sets.

(cont.) Third, we develop an algorithm for the retrospective topology correction of 3D digital segmentations. Our method is nested in the theory of Bayesian parameter estimation, and integrates statistical information into the topology correction process. In addition, no assumption is made on the topology of the initial input images. Finally, we propose a genetic algorithm to accurately correct the spherical topology of cortical surfaces. Unlike existing approaches, our method is able to generate several potential topological corrections and to select the maximum-a-posteriori retessellation in a Bayesian framework. Our approach integrates statistical, geometrical, and shape information into the correction process, providing optimal solutions relatively to the MRI intensity profile and the expected curvature.