Classifying Papilledema Using Machine Learning – Modernizing Frisén Grading
Joseph Branco1, Juo-Kai Wang2, Tobias Elze4, Mona Garvin3, David Szanto5, Randy Kardon6, Louis Pasquale1, Mark Kupersmith1
1Icahn School of Medicine At Mount Sinai, 2Ophthalmology, University of Iowa School of Medicine, 3University of Iowa School of Medicine, 4Harvard University, 5SUNY Medical School at Stony Brook, 6Univ of Iowa Hosp and Clinics
Objective:
We hypothesize that machine learning (ML) can quantify papilledema and detect a treatment effect on papilledema due to idiopathic intracranial hypertension (IIH).
Background:
ML can differentiate papilledema from normal optic discs. Currently, papilledema severity is assessed using the descriptive, ordinal Frisén scale.
Design/Methods:
We trained a convolutional neural network (CNN) to autonomously assign a Frisén grade using 2608 fundus photos from fellow eyes of 158 participants in the IIH Treatment Trial (IIHTT) and both eyes from eight clinic patients with grade 4 or 5 papilledema. Experts in classifying papilledema previously graded the photos. Our validation set consisted of 2969 photos of the IIHTT study eyes (for each participant, the eye with worse vision). To investigate the change over time, we divided study eyes into treatment groups, acetazolamide + diet (ACZ) vs. placebo + diet.
Results:
ML produced continuous values from 0-5, which identified photos that contained features from more than one Frisén grade. Activation maps showed the model focused on the optic disc. The average predicted Frisén grade correlated strongly with ground truth (r = 0.80, p < 0.001; mean absolute error = 0.49). At presentation, treatment groups had similar ML Frisén grades. The average ML Frisén grade for the ACZ treatment group (1.6, 95% CI 1.5-1.8) was significantly lower (p < 0.01) than for the placebo group (2.2, 95% CI 1.9-2.5) at the six month trial outcome. This difference was noted as early as one month (p = 0.03).
Conclusions:
Supervised ML of fundus photos successfully grades the degree of papilledema and tracks the changes, reflecting the effects of ACZ therapy. Given the increasing availability of fundus photography, neurologists will be able to utilize ML to quantify papilledema on a continuous scale that incorporates the descriptive features of the Frisén grade to monitor treatment of papilledema.