Semi-supervised Learning for Nerve Segmentation in Corneal Confocal Microscope Photography

Abstract

Corneal nerve fiber medical indicators are promising metrics for diagnosis of diabetic peripheral neuropathy. However, automatic nerve segmentation still faces the issues of insufficient data and expensive annotations. We propose a semi-supervised learning framework for CCM image segmentation. It includes self-supervised pre-training, supervised fine-tuning and self-training. The contrastive learning for pre-training pays more attention to global features and ignores local semantics, which is not friendly to the downstream segmentation task. Consequently, we adopt pre-training using masked image modeling as a proxy task on unlabeled images. After supervised fine-tuning, self-training is employed to make full use of unlabeled data. Experimental results show that our proposed method is effective and better than the supervised learning using nerve annotations with three-pixel-width dilation.