Automatic prostate segmentation using multiobjective active appearance model in MR images

Abstract

Prostate cancer is the second largest cause of mortality among men. Prostate segmentation, i.e. theprecise determination of the prostate region in magnetic resonance imaging (MRI), is generally used for prostate volumemeasurement, which can be used as a potential prostate cancer indicator. This paper presents a new fully automaticstatistical model called the multiobjective active appearance model (MOAAM) for prostate segmentation in MR images.First, in the training stage, the appearance model, including the shape and texture model, is developed by applyingprincipal component analysis to the training images, already outlined by a physician. Then noise and roughness areproperly removed in the preprocessing step by Sticks filter and nonlinear filtering. This helps us provide the properconditions for the prostate region detection. Finally, in order to detect the prostate region, a new multiobjectivefunction is optimized using a suitable search algorithm. The proposed method has been applied to prostate images forsegmenting the prostate boundaries. The evaluation results indicate that the presented method can yield a DSC value of87.4 ± 5.00%, is less sensitive to the edge information and initialization, and has a stronger capture range in comparisonwith existing methods.