Statistical significance based graph cut regularization for medical image segmentation

摘要

Graph cut minimization formulates the image segmentation asa linear combination of problem constraints. The salient constraintsof the computer vision problems are data and smoothness which arecombined through a regularization parameter. The main task of theregularization parameter is to determine the weight of the smoothnessconstraint on the graph energy. However, the difference in functionalforms of the constraints forces the regularization weight to balancethe inharmonious relationship between the constraints. This paperproposes a new idea: bringing the data and smoothness terms on thecommon base decreases the difference between the constraint functions.Therefore the regularization weight regularizes the relationshipbetween the constraints more effectively. Bringing the constraints onthe common base is carried through the statistical significancemeasurement. We measure the statistical significance of each term byevaluating the terms according to the other graph terms. Evaluatingeach term on its own distribution and expressing the cost by the samemeasurement unit decrease the scale and distribution differencesbetween the constraints and bring the constraint terms on similarbase. Therefore, the tradeoff between the terms would be properlyregularized. Naturally, the minimization algorithm produces bettersegmentation results. We demonstrated the effectiveness of theproposed approach on medical images. Experimental results revealedthat the proposed idea regularizes the energy terms more effectivelyand improves the segmentation results significantly.