Image segmentation for highly variable anatomy : applications to congenital heart disease

Abstract

Automated segmentation of medical images can facilitate clinical tasks in diagnosis, patient monitoring, and surgical planning. However, current methods either rely on explicit correspondence detection, or use machine learning techniques that require a large collection of fully annotated and representative images. Neither of these approaches are suitable when anatomical variability is high and labeled data is limited. In this thesis, we formulate new interactive segmentation methods and evaluate their applicability to congenital heart disease, which involves a wide range of cardiac malformations and topological changes and for which few image analysis methods have been previously developed. We begin by describing the new imaging datasets that we have created to support our research in congenital heart disease. Next, we show that image patches can be used to exploit manual segmentations made on a small set of slice planes in order to automatically segment the rest of an image, and investigate the potential of active learning to automatically solicit user input. Third, we develop an iterative segmentation model that can be accurately learned from small datasets which do not necessarily include the same pathologies as a new image to be segmented, and demonstrate that our model better generalizes to patients with the most severe heart malformations. Ultimately, the methods developed here take a step towards bringing the benefits of medical image analysis to challenging clinical applications involving large anatomical variability and small datasets.