Comparison of ASM and CNN based prostate segmentation in CT images

Abstract

Accurate localisation of the prostate boundaries in the CT data is a necessary step in radiation therapy treatment planning. Manual outlining is a challenging real-world problem, thus an automated methods are needed to relieve the medical doctors. This, however, is not a trivial task due to unclear boarder between the neighbouring organs and variation in shape, size and location of the prostate itself. Recent advances in deep learning show applications in semantic image segmentation having performances superior to the traditional, non learning based methods. In this paper, we compare performance of the two approaches: one based on the Active Shape Models and the second exploiting the convolutional neural networks (CNN). Both compared methods were trained using CT volumes belonging to the same set of the real patients data annotated by clinical experts. CNN turned out to offer better performances in terms of accuracy and robustness. Segmentation accuracy reached by both methods was measured using Dice Similarity Coefficient (DSC) and Jacard Coefficient (JC) and was equal in average 0,720 DSC, 0.598 JC, and 0.796 DSC, 0.663 JC for ASM and CNN respectively.