TH‐CD‐206‐06: Regularized Composite Shape Prior Encoding Shape Relevance in Variational Image Segmentation

Abstract

Purpose:Variational image segmentation incorporating regularized composite shape prior (RCSP) based on a shape dictionary has demonstrated benefits in robustness and accuracy. However, it still has the risk of local minimum in a non‐convex optimization setting. This study aims to drive the variational segmentation towards global optimum by introducing hyper prior on the composite weights of RCSP, which encode the shape priors’ relevance to the specific segmentation task.Methods:In addition to using a RCSP to regularize the variational segmentation, where the RCSP is constructed by a linear combination of the shape dictionary, this study introduces a hyper prior on the linear weights of this shape composite. More specifically, geometric relevance of each shape in the dictionary to the unknown target segmentation is inferred from image based surrogate metrics. Such relevance value is used as hyper prior and imposed on the linear weights of the composite shapes. The resulted RCSP with hyper prior regularization is incorporated in a unified active contour optimization framework and a variational block‐descent algorithm is derived and implemented.Results:The performance was assessed on corpus callosum segmentation using a brain MR dataset, and compared with typical benchmark approaches. The resulted RCSP demonstrated proper composition of training data w.r.t. their individual geometric relevance. The accuracy of ultimate segmentation estimates yielded statistically significant improvement with mean and medium Dice similarity coefficient (DSC) of (.946, .963) compared to (.902, .919), (.932, .946), and (.944, .964) for ATLAS based scheme, Chan Vese active contour model and an existing RCSP model, respectively.Conclusion:This work has developed a hyper prior to encode shape priors’ geometric relevance for RCSP regularization in a variational segmentation framework, leading to superior segmentation over benchmark approaches.