Atlas diffeomorphisms via object models

To tackle the problem of segmenting several closely-spaced objects from 3D medical images, I propose a hybrid of two segmentation approaches: one image-based and one model-based. A major contribution takes the image-based approach by diffeomorphically mapping a fully segmented atlas image to a partially segmented target patient image preserving any 'correspondence' inferred from the partial segmentation of the target. The mapping is produced by solving the steady-state heat flow equation where the temperature is a coordinate vector and corresponding points have the same temperature. Objects carried over from the atlas into the target serve as reasonable initial segmentations and can be further refined by a model-based segmentation method. Good quality segmentations are added to the list of the initial partial segmentations, and the process is repeated. Another contribution takes the model-based approach in developing shape models of quasi-tubular objects and statistics on those models. Whereas medial models were previously only developed for slab-shaped objects, this contribution provides an approximately medial method to stably represent nearly tubular objects. I test my method on segmenting objects from 3D Computed Tomography (CT) scans of the head and neck obtained for radiotherapy treatment planning.