Cataract Diagnosis at Scale: Leveraging Cloud Resources for Accurate and Eco-Friendly Fundus Image Analysis with AI

Abstract

Purpose : This study presents the development and evaluation of an artificial intelligence (AI) model designed to distinguish fundus images of individuals with cataracts from those without. The primary objective is to contribute to the advancement of automated diagnostic tools for ocular health assessment.

Methods : The AI model was trained on Google's Collaboration platform using 2,074 fundus images sourced from a public database on Kaggle.com. Leveraging the platform's computational resources, the model underwent training for 8 node hours, resulting in a cost-free and carbon-neutral process. Google's servers facilitated efficient model training and eliminated associated computational expenses. The dataset was meticulously curated to include a diverse set of images, ensuring comprehensive coverage of cataract-related variations. The model's performance was assessed using standard metrics, including accuracy, precision, recall (sensitivity), specificity, and F1-score.

Results : The developed AI model exhibited commendable diagnostic accuracy, achieving an overall accuracy of 96.5%. Precision, indicating the model's ability to correctly identify cataract-positive cases, was measured at 97.9%. A high recall rate of 95% underscored the model's effectiveness in capturing true cataract instances. Specificity, gauging the model's aptitude for correctly recognizing cataract-free cases, was reported at 98%. The F1-score, representing a harmonized balance between precision and recall, stood at an impressive 96.4%. These robust performance metrics position the AI model as a promising tool for cataract detection in fundus images.

Conclusions : This study showcases the successful development of an AI model capable of discerning between fundus images of individuals with cataracts and those without. The utilization of Google's Collaboration platform not only facilitated cost-effective and environmentally friendly model training but also demonstrated the potential for leveraging cloud-based resources in medical image analysis. The high accuracy, precision, recall, specificity, and F1-score affirm the efficacy of the proposed model, suggesting its utility as a valuable asset in automated cataract diagnosis and ocular health assessment.

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

 

Confusion matrix demonstrating the performance of the model.

 

Area-under-the-curve (AUC) graph depicting the recall and precision of the model (confidence interval set to .05)