Automated Segmentation of Prostate MR Images Using Prior Knowledge Enhanced Random Walker

Prostate cancer is the second most common cause of cancer deaths in males. Accurate prostate segmentation from magnetic resonance (MR) images is critical to the diagnosis and treatment of prostate cancer. Automated prostate segmentation is challenging due to the variety in shapes and sizes of the prostate. Furthermore, the expected boundaries of ROIs are often indistinct, while heterogeneity concurrently exists within the ROIs. To address these challenges, we propose an automated approach that incorporates the local intensity features by random walker (RW) algorithm and global probability knowledge from an atlas to better describe unique characteristics of the prostate in MR images. We formulated a new RW weight function to take into account atlas probabilities and intensity differences. The prior knowledge from the atlas probability map not only reflects the statistical shape approximation of the prostate but also provides confinement and guidance for RW segmentation. Our approach was validated and compared with the conventional RW algorithm on segmenting 30 3-T prostate MR volumes. The experimental results indicated that our approach with an average DSC of 80.7±5.1%, outperformed that of the conventional RW (average DSC = 71.9±9.1%) and several other reported methods in terms of DSC accuracy and robustness.