Weakly Supervised Segmentation for Building Statistical Models

Objective:

The goal of this project is to investigate and develop unsupervised or minimally supervised techniques for the segmentation of the anatomy (in particular, bone structures).

Statistical models of variability in anatomical structures are essential for a wide range of medical applications, e.g., to reconstruct missing information in implant design or help localize bones for radiotherapy planning. Segmentation of the anatomy is a necessary step towards such robust statistical shape and intensity modeling. A major obstacle in generating such statistical models is the need for a sufficiently large amount of correctly segmented medical data, from which the statistical model can be derived. Besides the purposes of active shape modelling based segmentation, it is beneficial to develop prior-less or weakly-supervised techniques to prepare the initial data used for modelling, and to deal with cases not well described by the statistical shape model (e.g. trauma, pathology).

Conventional segmentation methods for such large sets on segmentation data require either a significant amount of manual, interactive effort of the medical personnel or a statistical shape model of the bone in question, and sometimes a combination of both. The objective of this project is to develop weakly supervised methods for segmentation for subsequent automatic generation of statistical models. To this end, we are utilizing single atlas methods to provide semantic context in a general segmentation framework. We are particularly interested in bone segmentation to be utilized in the generation of a statistical skeleton atlas of the body.

Participants: Prof. Orçun Göksel, Dr. Tobias Gass, Prof. Gábor Székely