Probabilistic Refinement of Model-Based Segmentation: Application to Radiation Therapy Planning of the Head and Neck

Abstract

Radiation therapy planning requires accurate delineation of target volumes and organs at risk. Traditional manual delineation is tedious, and can require hours of clinician’s time. The majority of the published automated methods belong to model-based, atlas-based or hybrid segmentation approaches. One substantial limitation of model-based segmentation is that its accuracy may be restricted either by the uncertainties in image content or by the intrinsic properties of the model itself, such as prior shape constraints. In this paper, we propose a novel approach aimed at probabilistic refinement of segmentations obtained using 3D deformable models. The method is applied as the last step of a fully automated segmentation framework consisting of automatic initialization of the models in the patient image and their adaptation to the anatomical structures of interest. Performance of the method is compared to the conventional model-based scheme by segmentation of three important organs at risk in the head and neck region: mandible, brainstem, and parotid glands. The resulting segmentations are validated by comparing them to manual expert delineations. We demonstrate that the proposed refinement method leads to a significant improvement of segmentation accuracy, resulting in up to 13% overlap increase.