Bone fragment segmentation from 3D CT imagery using the Probabilistic Watershed Transform

This paper presents a novel method to segment images and applies this method for segmenting bone fragments imaged using 3D Computed Tomography (CT). Existing image segmentation solutions tend to have difficulty in accurately delineating regions that have subtle variations along their boundaries or delineating regions which are spatially close. The proposed image segmentation algorithm introduces an original modification to the classical watershed transform and we refer to resulting approach as the Probabilistic Watershed Transform (PWT). The PWT uses a set of probability distributions to model the likelihood that a given pixel is a measurement obtained from each of the provided semantic classes. While the framework for the proposed PWT allows for completely general likelihood distributions, we specify several likelihood distributions which address known shortcomings in the watershed transform and, more generally, competing segmentation methods. Using these likelihood distributions, we apply the PWT to segment bone fragments within CT images of a bone fracture. A quantitative evaluation of the bone segmentation results is provided which compares our results with several leading competing methods as well as human-generated segmentation which show that the proposed method has some significant benefits for solving the bone fragment segmentation problem.