A semiautomatic segmentation approach to corneal lesions

The cornea is an essential structure for the proper functioning of human vision. It can suffer injuries like tumors, areas of epithelial removal, infections, and post-surgical injuries that need prompt and effective treatment. The affected region’s area evolution monitoring enables the physician to evaluate the treatment effectiveness. This paper presents a semi-automatic method that can assist the physician in monitoring the evolution of corneal lesions. Our approach uses some specialist-marked regions to train a random forest classifier, and then to classify the other image areas as lesion or non-lesion. We extract color information, and, after classification, we apply an active contour operation to the most significant connected component. Our tests show that by marking 5% of the pixels, our method achieves an accuracy of 99.08% and a Dice of 0.85 on average. According to the literature, the segmentation in more than 90% of the images was considered excellent.