IMPROVED SEGMENTATION AND DEFORMITY CORRECTION IN MEDICAL IMAGING SYSTEMS

Abstract

Medical image segmentation plays an important role in computer-aided diagnosis by providing exact marking of or tracing of anatomical structures. In this domain, although deep convolutional neural networks have demonstrated impressive performances, the segmented outputs usually do not show the value of stability and accuracy that are required for clinical deployment. We present in this article the novel SESV (Segmentation Error-based Segmentation and Verification) framework, aimed at further improving the segmentation performance of pre-existing DCNN models while leaving their basic architecture intact. Instead of attempting to improve segmentation accuracy alone, SESV identifies and corrects segmentation errors by first producing error maps which serve as priors. The error map, the original image, and the initial segmentation mask then enter the re-segmentation network, which looks to derive a refined segmentation. In order to make sure, the verification network is set up to do verification and decide whether it is proper to accept the corrected segmentation. Experimental results show that the SESV framework considerably boosted the quality of the segmentation and offered a sound alternative toward medical