Prostate lesion segmentation in MR images using radiomics based deeply supervised U-Net

Abstract

Prostate lesion detection in an axial T2 weighted (T2W) MR images is a very challenging task due to heterogeneous and inconsistent pixel representation surrounding the prostate boundary. In this paper, a radiomics based deeply supervised U-Net is proposed for both prostate gland and prostate lesion segmentation. The proposed pipeline is trained and validated on 1174 and 2071 T2W MR images of 40 patients and tested on 250 and 415 T2W MR images of 10 patients for prostate capsule segmentation and prostate lesion segmentation, respectively. Effective segmentation of prostate lesions in various stages of prostate cancer (namely T1, T2, T3, and T4) is achieved using the proposed framework. The mean Dice Similarity Coefficient (DSC) for actual prostate capsule segmentation and prostate lesion segmentation is 0.8958 and 0.9176, respectively. The proposed framework is also tested on Promise12 public dataset for performance analysis in segmenting prostate gland. The segmentation results using proposed architecture are promising compared to state-of-the-art techniques. It also improves the accuracy of the prostate cancer diagnosis.