A Computational Method for Image Segmentation Based on a Dirichlet Field and an Analysis of the Asymptotic Accuracy of Spatial Regularizer Discretization

This paper presents a Dirichlet-field formulation of segmentation with closed-form uncertainty estimation, edge-aware smoothing, and an explicit asymptotic analysis of the spatial regularizers used in the model.

Numerical experiments on ACDC, Synapse, and CHAOS show improved calibration and segmentation quality with moderate extra cost.