AI-Assisted Diagnosis of Corneal Ulcers: A Sustainable and Precise Approach for Improving Clinical Decision Making and Patient Outcomes

Abstract

Purpose : Corneal ulcers are defects in the corneal epithelium that can potentially lead to significant visual impairment if not promptly treated. It is important to note that urgency of treatment will also depend on the severity of the condition. Corneal ulcers can be classified into several categories according to their shape and distribution—from relatively mild point-like corneal ulcers to more serious and advanced flaky corneal ulcers (Deng et al., 2020). Our primary goal is to explore grounds for supplementing clinical diagnoses with artificial intelligence by developing a precise and efficient tool to aid ophthalmologists in promptly identifying the categorical grade of corneal ulcers.

Methods : Using a dataset of 712 publicly available images obtained from Kaggle.com, our AI model went through extensive training on Google's Collaboration platform for a duration of 2 hours and 23 minutes. The training process was not efficient but also environmentally friendly with no associated costs or carbon footprint. The dataset consisted of slit lamp images showcasing ulcers in both point like and point flaky mixed forms. We evaluated the performance of the AI model using metrics such as accuracy, precision, recall (sensitivity) specificity and F1 score.

Results : The AI model showed performance results achieving an accuracy rate of 88.5% a precision score of 88.7%, a recall rate of 88.7% a specificity level of 90.5% and an F1 score of 88.8%. Additionally, the model demonstrated an outstanding Area Under the Curve (AUC) value of 0.982, signifying its robust discriminatory power in distinguishing between categorical corneal ulcer types. These outcomes demonstrate the model's capability to accurately differentiate forms of corneal ulcers highlighting its potential for practical use, in clinical settings.

Conclusions : The research showcases an AI model that can distinguish between two types of corneal ulcers, namely point like and point flaky in slit lamp images. The use of Google's Collaboration platform enabled cost effective training of the model presenting a sustainable approach to advancing medical diagnostics with artificial intelligence. The impressive accuracy, AUC, precision, recall, specificity and F1 score achieved highlight the potential of this model as a tool for enhancing clinical decision making and improving patient outcomes in diagnosing corneal ulcers.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.