Bayesian Image Segmentation with Multiscale Feature Learning

PhD-defense by Kersten Petersen


This thesis is concerned with developing a Bayesian image segmentation method that models local appearance information with multiscale feature learning techniques. The proposed approach is particularly suited for real-world segmentation tasks, where complex appearance information needs to be combined with prior knowledge. The main constituents of the presented Bayesian framework are a local appearance model, a shape prior, and a location prior. Instead of handcrafting features for segmentation, we introduce a local appearance model that is able to learn informative features directly from the data. Our method achieves state-of-the-art performance on segmenting lumbar vertebrae and two real-world benchmark datasets from computer vision.

A static sequential Monte Carlo (SMC) sampler on shapes is proposed to perform efficient inference in the Bayesian framework. Interactive segmentation is supported by incorporating a novel conditional version of the Point Distribution Model from Cootes and Taylor. The presented segmentation method is applied to automatically score the cardiovascular risk (CVD) from standard lumbar radiographs. To achieve this goal we segment abdominal aortic calcifications, which have been shown to significantly correlate with the event of cardiovascular death. It is a challenging segmentation problem, because the structures of interest are barely discernible and require large contextual regions to be discriminated from clutter and other anatomical structures.

Assessment Committee:

Chairman: Associate Professor Marleen de Bruijne, Department of Computer Science, Copenhagen University

Member 1: Professor Milan Sunka, Dept. of Electrical and Computer Engineering, University of Iowa

Member 2: Professor Tim Cootes, Human and Medical Sciences, University of Manchester

Academic supervisor: Mads Nielsen

For an electronic copy of the thesis, please contact Dina Riis Egholm -