Automatic Screening for Ocular Anomalies Using Fundus Photographs.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Ocular anomaly screening via fundus photography is crucial for preventing vision impairment.
• Automated algorithms are needed to manage increasing screening demands and expert shortages.

Purpose of the Study:

• Develop a deep learning algorithm for detecting ocular anomalies in fundus photographs.
• Evaluate the algorithm's performance in classifying "normal versus anomalous" eye examinations in diabetic and general populations.

Main Methods:

• A deep learning algorithm was developed and trained on large datasets from diabetic (OPHDIAT) and general (OphtaMaine) patient cohorts.
• The algorithm was fine-tuned using a general population dataset for improved performance.
• Algorithm performance was compared against ophthalmologist diagnoses on a subset of examinations.

Main Results:

• The algorithm achieved high accuracy, with an area under the ROC curve of 0.9592 on the diabetic test set.
• Fine-tuning significantly improved performance on the general population test set (AUC 0.9108).
• The fine-tuned algorithm demonstrated superior sensitivity (0.8248) compared to human performance (0.6682) at a similar specificity.

Conclusions:

• The developed deep learning algorithm effectively classifies normal versus anomalous eye examinations using fundus photography.
• This AI tool shows performance comparable to, and in some aspects exceeding, human experts.
• Comprehensive screening for ocular anomalies using AI is feasible, extending beyond specific retinal pathologies.