Deep learning automated detection of reticular pseudodrusen from fundus autofluorescence images or color fundus photographs in AREDS2.

Researchers

Journal

Modalities

Models

Abstract

To develop and evaluate deep learning models for detecting reticular pseudodrusen (RPD) using fundus autofluorescence (FAF) images or, alternatively, color fundus photographs (CFP), in the context of age-related macular degeneration (AMD).
Application of deep learning models to the Age-Related Eye Disease Study 2 (AREDS2) dataset.
11,535 FAF and 11,535 CFP images from longitudinal follow-up of 2,450 participants in the AREDS2 dataset. Gold standard labels were derived from human expert reading center grading of the FAF images and transferred to the corresponding CFP images.
A deep learning model was trained to detect RPD in eyes with intermediate to late AMD, using FAF images (FAF model). Using label transfer from FAF to corresponding CFP images, a second deep learning model was trained to detect RPD from CFP (CFP model). Model performance was compared with that of four ophthalmologists using a random subset from the full test set.
Area under the curve (AUC); kappa; accuracy; F1-score.
On the full test set, the FAF model had AUC 0.939 (95% confidence interval 0.927-0.950), kappa 0.718 (0.685-0.751), accuracy 0.899 (0.887-0.911), and F1-score 0.783 (0.755-0.809). The CFP model had equivalent values of 0.832 (0.812-0.851), 0.470 (0.426-0.511), 0.809 (0.793-0.825), 0.593 (0.557-0.627), respectively. The FAF model demonstrated superior performance to four ophthalmologists on the random subset, showing higher kappa of 0.789 (0.675-0.875) versus range 0.367-0.756, higher accuracy 0.937 (0.907-0.963) versus range 0.696-0.933, and higher F1-score 0.828 (0.725-0.898) versus range 0.539-0.795. The CFP model demonstrated substantially superior performance to four ophthalmologists on the random subset, showing higher kappa 0.471 (0.330-0.606) versus range 0.105-0.180, higher accuracy 0.844 (0.798-0.886) versus range 0.717-0.814, and higher F1-score 0.565 (0.434-0.674) versus range 0.217-0.314.
Deep learning-enabled automated detection of RPD presence from FAF images achieved a high level of accuracy, equal or superior to that of ophthalmologists. Automated RPD detection using CFP achieved a lower accuracy that still surpassed that of ophthalmologists. Deep learning models can assist, and even augment, the detection of this clinically important, AMD-associated lesion.
Copyright © 2020. Published by Elsevier Inc.