Options
2016
Conference Paper
Titel
Statistical shape modeling from gaussian distributed incomplete data for image segmentation
Abstract
Statistical shape models are widely used in medical image segmentation. However, getting sufficient high quality manually generated ground truth data to generate such models is often not possible due to time constraints of clinical experts. In this work, a method for automatically constructing statistical shape models from incomplete data is proposed. The incomplete data is assumed to be the result of any segmentation algorithm or may originate from other sources, e.g. non expert manual delineations. The proposed work flow consists of (1) identifying areas of high probability in the segmentation output of being a boundary, (2) interpolating between the boundary areas, (3) reconstructing the missing high frequency data in the interpolated areas by an iterative back-projection from other data sets of the same population. For evaluation, statistical shape models where constructed from 63 clinical CT data sets using ground truth data, artificial incomplete data, and incomplete data resulting from an existing segmentation algorithm. The results show that a statistical shape model from incomplete data can be built with an added average error of 6 mm compared to a model built from ground truth data.