A Coarse to Fine Corneal Ulcer Segmentation Approach Using U-net and DexiNed in Chain

Abstract

A corneal ulcer is one of the most frequently appearing diseases that may affect eye health. The proper measurement of corneal ulcer lesions enables the physician to evaluate the treatment effectiveness and assist in decision-making. This article presents the solution for ulcer segmentation as a pixel-wise classification task, and proposes a novel coarse-to-fine method to extract corneal ulcers from ocular staining images. This study combines two classical convolutional neural networks (CNNs), known as U-net and DexiNed, following Morphological Geodesic Active Contour as a post-processing operation. We trained the CNNs using 358 point-flaky corneal ulcer images and evaluated its performance in 91 flaky corneal ulcer images. Our approach achieved 70.50% of Dice Coefficient on average, 87.4% of Recall, and 99.0% of Specificity, and True Dice Coefficient of 63.7%. These results corroborate our approachs efficacy and efficiency.