eduzhai > Applied Sciences > Engineering >

Cross-denoising Network against Corrupted Labels in Medical Image Segmentation with Domain Shift

  • Save

... pages left unread,continue reading

Document pages: 7 pages

Abstract: Deep convolutional neural networks (DCNNs) have contributed manybreakthroughs in segmentation tasks, especially in the field of medicalimaging. However, textit{domain shift} and textit{corrupted annotations},which are two common problems in medical imaging, dramatically degrade theperformance of DCNNs in practice. In this paper, we propose a novel robustcross-denoising framework using two peer networks to address domain shift andcorrupted label problems with a peer-review strategy. Specifically, eachnetwork performs as a mentor, mutually supervised to learn from reliablesamples selected by the peer network to combat with corrupted labels. Inaddition, a noise-tolerant loss is proposed to encourage the network to capturethe key location and filter the discrepancy under various noise-contaminantlabels. To further reduce the accumulated error, we introduce aclass-imbalanced cross learning using most confident predictions at theclass-level. Experimental results on REFUGE and Drishti-GS datasets for opticdisc (OD) and optic cup (OC) segmentation demonstrate the superior performanceof our proposed approach to the state-of-the-art methods.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×