Radiomics for peripheral zone and intra-prostatic urethra segmentation in MR imaging

Abstract Automatic peripheral zone (PZ) and intra-prostatic urethra segmentation has clinical significance in analysis of prostate health management. It is interesting and much challenging task due to heterogeneous and inconsistent pixel intensities around prostate boundary and changes in a shape of the actual prostate capsule from patient to patient. The traditional methods of detection and delineation for glandular prostate gland using magnetic resonance imaging (MRI) involve expertise of radiologists and it is expensive in terms of time and accuracy. This paper proposes a novel technique for automate segmentation of PZ of prostate and intra-prostatic urethra. The technique is based on radiomics extraction using nonnegative matrix factorization (NMF) and segmentation using the self organizing maps (SOMs). The proposed framework is evaluated using 52 axial T2 weighted (T2w) MR images. The dice similarity coefficient (DSC) is calculated to measure the similarity between segmentation results and ground truth images. The proposed algorithm is compared with the conventional K-means (KM) clustering and fuzzy C-means (FCM) clustering approaches of the segmentation. The proposed scheme is shown to be superior based on subjective and objective evaluation analysis. The average percentage of DSC for PZ and intra-prostatic urethra segmentation is 87.33% and 85.55%, respectively using the proposed technique.