Computed Aided Prostate Segmentation of Multiparametric MR Images Using Latent Features

Abstract

This paper proposes a computer aided prostate segmentation system from Multiparametric Magnetic Resonance Imaging (MRI). Computer assisted segmentation of prostate from MRI is a fundamental task. The dim prostate boundary and large diversity in prostate shape among the patients are the main challenges in automatic prostate segmentation. To deal with these challenges, we used image patches to build likelihood map. An energy function is minimized to identify the patches as the prostate or non-prostate region by using Atlas based segmentation. The hidden features were derived from prostate MRI by deploying auto encoders. These learned features are more effective and compact in comparison to other features to describe undisclosed data. These refined features are processed to derive likelihood map for unknown test MR image. Active contour model is also deployed to perform final segmentation. The proposed computer aided detection system is analyzed effectively on T2-weighted prostate MR images of 37 patients. Hausdorff Distance (HD) and Dice Similarity Coefficient (DSC) are used to assess the presented system by considering ground-truth prepared by skilled radiologist. The Dice coefficient obtained by proposed system is of 91.47% ± 4.67%, and HD of 7.46 mm ±2.67 mm and classifier accuracy of 88.1% in cancerous tissue detection. The initial results manifest that the hidden features are adequate in prostate segmentation from MRI.

Keywords– Prostate, Segmentation, Magnetic resonance imaging (MRI), Deformable model, Prostate cancer, Deep learning