Classifying and quantifying changes in papilloedema using machine learning

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Papilledema, a sign of increased intracranial pressure, is currently graded using the subjective Frisén scale.
• Machine learning (ML) shows potential for objective assessment of papilledema using fundus photography.
• This study investigates ML's ability to quantify papilledema and treatment effects in idiopathic intracranial hypertension (IIH).

Purpose of the Study:

• To develop and validate a ML model for grading papilledema severity from fundus images.
• To assess if ML can detect treatment effects on papilledema in IIH patients.
• To compare ML-based grading with the established Frisén scale.

Main Methods:

• A convolutional neural network was trained to assign Frisén grades to fundus images from the IIHTT.
• A fivefold cross-validation approach was used on 2979 images from 158 participants.
• ML-based grades were compared to expert grades and analyzed for treatment group differences.

Main Results:

• ML-determined grades showed strong correlation with expert grades (r=0.76).
• The ML model achieved a mean absolute error of 0.54.
• At 6 months, the acetazolamide (ACZ) group showed significantly lower ML-graded papilledema (1.7) compared to the placebo group (2.3).

Conclusions:

• Supervised ML effectively quantifies papilledema degree and temporal changes from fundus photos.
• ML provides a continuous scale for papilledema, incorporating Frisén scale features.
• This technology can aid neurologists in monitoring interventions for IIH.