Abstract:

Medical image segmentation is a crucial step of computer-aided diagnosis. Although DCNN has achieved great success in such a task, the resulting segmented images are not accurate and stable enough for clinical application. In this work, rather than trying to improve segmentation accuracy, we introduce a novel SESV framework that boosts up the accuracy of current available DCNNs for performing medical image segmentation. It takes its stand by prediction and correction of the error produced due to segmentation created by the use of the given model. Errors in classification cannot be foretold straight away due to some unavoidable challenges involved; thus we have presented a strategy called segmentation faults: Using error maps first as priors instead of using the masks created from the corrected parts directly. This error map, together with the original image and segmentation mask, is passed through a re-segmentation network. Finally, we propose a verification network to decide whether the corrected mask from the re-segmentation process should be accepted.

Description:

.