Dilation-Assisted Contrastive Learning for Corneal Nerve Segmentation in Corneal Confocal Microscopy

Accurate extraction of corneal nerves in corneal confocal microscopy images is essential for assessing morphological changes linked to various eye-related diseases. However, current segmentation methods struggle with distinguishing corneal nerves from visually similar structures. In this paper, we propose a dilation-assisted contrastive learning (Dicon) model. The proposed Dicon incorporates a dilation-guided positive pairing mechanism that leverages structural priors from ground truth to guide the model toward more precise extraction of corneal nerves. Additionally, we propose a selective feature amplification regularization strategy, which uses dilation masks to differentiate between foreground and background, thereby enhancing the model's ability to distinguish between visually similar structures while reinforcing the learning of corneal nerve features. Experimental results demonstrate that Dicon improves segmentation sensitivity by at least 2% compared to state-of-the-art methods.