Caudate Segmentation using Deformable M-reps

We use object scale and then atom scale Bayesian optimization of m-reps to automatically segment the caudate nucleus in brain MRI images. Our shape priors are learned after alignment of m-reps fit to 15 manual segmentations of caudates. At the object and atom scale levels the alignment is to the m-rep mean of the object and atom, respectively. Our appearance likelihood is learned from regional intensity quantile functions from images that have been aligned, corrected for inhomogeneity, and intensity normalized. We begin the segmentation of a target image by the image preprocessing steps described above followed by an initialization of the mean m-rep model using the image alignment transformation. The segmentation then proceeds by optimizing the posterior probability over a shape space encompassing eigenmodes of full m-rep shape variation capturing 87% of the total variance in the training population of left caudates (83% right). The segmentation concludes by successively optimizing the posterior probability of the residual changes in the medial atoms making up the m-rep. Since only a weak variant of a part of our segmentation method was inadvertently applied in the Grand Challenge, its results are not representative of the performance of our method. Qualitative results of the actual method are reported within; quantitative results will be reported separately.