Label Set Perturbation for MRF based Neuroimaging Segmentation

Graph-cuts based algorithms are effective for a variety of segmentation tasks in computer vision. Ongoing research is focused toward making the algorithms even more general, as well as to better understand their behavior with respect to issues such as the choice of the weighting function and sensitivity to placement of seeds. In this paper, we investigate in the context of neuroimaging segmentation, the sensitivity/stability of the solution with respect to the input "labels" or "seeds". In particular, as a form of parameter learning, we are interested in the effect of allowing the given set of labels (and consequently, the response/ statistics of the weighting function) to vary for obtaining lower energy segmentation solutions. This perturbation leads to a "refined" label set (or parameters) better suited to the input image, yielding segmentations that are less sensitive to the set of labels or seeds provided. Our proposed algorithm (using Parametric Pseudoflow) yields improvements over graph-cuts based segmentation with a fixed set of labels. We present experiments on about 450 3-D brain image volumes demonstrating the efficacy of the algorithm.