Automatic segmentation of the prostate in 3D MR images by atlas matching using localized mutual information

Abstract

An automatic method for delineating the prostate (including the seminal vesicles) in three‐dimensional magnetic resonance scans is presented. The method is based on nonrigid registration of a set of prelabeled atlas images. Each atlas image is nonrigidly registered with the target patient image. Subsequently, the deformed atlas label images are fused to yield a single segmentation of the patient image. The proposed method is evaluated on 50 clinical scans, which were manually segmented by three experts. The Dice similarity coefficient (DSC) is used to quantify the overlap between the automatic and manual segmentations. We investigate the impact of several factors on the performance of the segmentation method. For the registration, two similarity measures are compared: Mutual information and a localized version of mutual information. The latter turns out to be superior (median , with a paired two‐sided Wilcoxon test) and comes at no added computational cost, thanks to the use of a novel stochastic optimization scheme. For the atlas fusion step we consider a majority voting rule and the “simultaneous truth and performance level estimation” algorithm, both with and without a preceding atlas selection stage. The differences between the various fusion methods appear to be small and mostly not statistically significant . To assess the influence of the atlas composition, two atlas sets are compared. The first set consists of 38 scans of healthy volunteers. The second set is constructed by a leave‐one‐out approach using the 50 clinical scans that are used for evaluation. The second atlas set gives substantially better performance , stressing the importance of a careful atlas definition. With the best settings, a median DSC of around 0.85 is achieved, which is close to the median interobserver DSC of 0.87. The segmentation quality is especially good at the prostate‐rectum interface, where the segmentation error remains below in 50% of the cases and below in 75% of the cases.