Detecting Inflammation in Fundus Photographs Using Machine Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Posterior segment inflammation diagnosis often relies on invasive imaging like ultra-widefield fluorescein angiography (UWFFA).
• Ultra-widefield fundus photography (UWFFP) offers a noninvasive imaging alternative.
• Developing AI tools to interpret UWFFPs for inflammation detection is crucial for accessible diagnostics.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model using UWFFPs as a surrogate for UWFFA in detecting posterior segment inflammation.
• To assess the diagnostic performance of ML-based UWFFP analysis compared to expert graders.

Main Methods:

• A dataset of 302 UWFFPs was curated, with UWFFA serving as the ground truth for inflammation classification.
• A single-label image classification model was trained using Vertex AI on UWFFPs to identify inflammation.
• The ML model's performance was compared against evaluations by fellowship-trained specialists and ophthalmologists on an independent set of UWFFPs.

Main Results:

• The ML model achieved an area under the curve of 0.943, with 90.91% sensitivity and 84.21% specificity for detecting inflammation.
• The model correctly diagnosed inflammation in 95% of additional UWFFPs, surpassing the accuracy of all human expert graders (85% to 65%).

Conclusions:

• UWFFPs, when analyzed with ML techniques, can serve as a noninvasive and accessible imaging modality for detecting posterior segment inflammation.
• AI-powered analysis of UWFFPs demonstrates superior or comparable accuracy to expert human interpretation for inflammation detection.
• This approach holds promise for improving the efficiency and accessibility of diagnosing posterior segment inflammatory conditions.