An automatic multi-atlas based prostate segmentation using local appearance-speci c atlases and patch-based voxel weighting

Prostate segmentation facilitates prostate cancer detection and can help to diagnose the pathological stage of disease. Segmented anatomical models may also help to improve the outcome of robotic-aided laparoscopic prostatectomy (RALP) by augmented reality image guidance. In this paper, we present a fully automated segmentation pipeline for multi-center and multi-vendor MRI prostate segmentation using a multi-atlas approach with local appearance-speci c voxel weighting. Segmenting prostates with a large variation of shape and intensity still remains a signi cant challenge. In this work, the atlases with the most similar global appearance are classi ed into the same categories. Sumof-square local intensity di erence after a ne registration is used for atlas selection and after non-rigid registration, a local patch-based atlas fusion is performed using voxel weighting based on the local patch distance. Such multi-atlas segmentation is a widely used method in brain segmentation. We thoroughly evaluated the method on 50 training images by performing a leave-one-out study. Dice coe cient and volumetric overlapping accuracy are used to quantify the di erence between the automatic and manual segmentation. Compared to the manual gold standard segmentation, our proposed method produced favorable outcomes in these highly variable data sets, with an average Dice coe cient 0.8467 ± 0.0435. The result shows that the algorithm presented could be used to aid the delineation of the prostate from diverse MRI images, which may be useful in a number of clinical applications.