PDCU-Net: A Depth Model by Imitating Expert Diagnostic Thinking for Corneal Ulcer Classification on Slit-Lamp Images

Infectious keratitis (IK) is a common and devastating ophthalmic disease that often leads to corneal ulcer. Timely and accurate initial diagnosis for corneal ulcer is required to ensure reasonable and accurate drug selection to prevent vision loss and even blindness. It is challenging for ophthalmologists to diagnose the type of infection and pattern of ulcer based on corneal ulcers. In this study, a Progressive Diagnostic Network for Corneal Ulcer Classification (PDCU-Net) is proposed. We introduce progressive training to imitate the thinking of ophthalmologists in diagnosis, and improve the model’s ability and interpretability to distinguish infection types and pattern of corneal ulcers through multi-stage progressive diagnosis. Furthermore, to enhance image details and increase the generalization ability of the model, we also design an image processing module. Experimental results on three datasets show that PDCU-Net achieves better performance than other state-ofthe-art classification networks. The model may accurately make a diagnosis of infection types and patterns of corneal ulcer from slit lamp images, which can effectively assist ophthalmologists in the objective and accurate diagnosis of corneal ulcer.