Utilizing human intelligence in artificial intelligence for detecting glaucomatous fundus images using...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma diagnosis relies on identifying subtle optic nerve and retinal changes.
• Traditional AI models often lack transparency, presenting a 'black box' challenge.
• Explainable AI (XAI) is crucial for clinical trust and adoption in medical diagnostics.

Purpose of the Study:

• To develop and validate a novel convolutional neural network (CNN) model for diagnosing glaucomatous damage using TrueColor confocal fundus images.
• To address the 'black box' problem in AI for medical diagnosis by incorporating human-in-the-loop (HITL) data annotation.
• To predict and precisely locate detailed glaucomatous signs, including splinter hemorrhages, optic atrophy, cupping, peripapillary atrophy, and RNFL defects.

Main Methods:

• A private dataset of 1,400 high-resolution confocal fundus images was curated, with 80% for training and 20% for testing.
• A custom You Only Look Once version 5 (YOLOv5) object detection model was employed for precise identification of conditions.
• Twenty-six medical conditions were annotated by a team of glaucoma specialists and optometrists using the Microsoft Visual Object Tagging Tool (VoTT).

Main Results:

• The AI model demonstrated consistent accuracy improvements during testing, reaching up to 98.89%.
• The model accurately predicted glaucoma diagnosis and identified detailed glaucomatous fundus signs.
• Testing involved three runs on separate image subsets (90, 100, 90 images) over 15-day intervals.

Conclusions:

• This study presents the first reported use of HITL machine learning for detecting glaucomatous fundus images.
• The developed AI model offers high sensitivity and specificity for glaucoma prediction.
• The model is an explainable AI, effectively overcoming the 'black box' dilemma in AI-driven diagnostics.