Deep learning on fundus images detects glaucoma beyond the optic disc

Abstract

Although unprecedented sensitivity and specificity values are reported, recent glaucoma detection deep learning models lack in decision transparency. Here, we propose a methodology that advances explainable deep learning in the field of glaucoma detection and vertical cup-disc ratio (VCDR), an important risk factor. We trained and evaluated deep learning models using fundus images that underwent a certain cropping policy. We defined the crop radius as a percentage of image size, centered on the optic nerve head (ONH), with an equidistant spaced range from 10-60% (ONH crop policy). The inverse of the cropping mask was also applied (periphery crop policy). Trained models using original images resulted in an area under the curve (AUC) of 0.94 [95% CI 0.92-0.96] for glaucoma detection, and a coefficient of determination (R²) equal to 77% [95% CI 0.77-0.79] for VCDR estimation. Models that were trained on images with absence of the ONH are still able to obtain significant performance (0.88 [95% CI 0.85-0.90] AUC for glaucoma detection and 37% [95% CI 0.35-0.40] R² score for VCDR estimation in the most extreme setup of 60% ONH crop). Our findings provide the first irrefutable evidence that deep learning can detect glaucoma from fundus image regions outside the ONH.

Figure 1

(A) R² values are plotted as a function of crop size, for both ONH crop and periphery crop policies. Evaluated crop sizes are indicated by data markers in the graph, and visualized in the two rows of fundus images. Results of occlusion are given as black triangles. (B) Close-up of example fundus image, with dotted lines corresponding to example crop sizes. (bottom panel) Detailed results for experiments following ONH cropping (C) and periphery cropping (D) policies. Kernel density estimation (KDE) plots with ground truth distribution on y-axis and prediction distribution on x-axis. The KDE plots also feature the MAE (top left), Pearson r (bottom right) and crop size (bottom left).

Figure 2

(A) AUC values are plotted as a function of crop size, for both ONH crop and periphery crop policies. Evaluated crop sizes are indicated by data markers in the graph. (B) Close-up of example fundus image (UZL), with dotted lines corresponding to example crop sizes. (bottom panel) Two rows of processed fundus images of REFUGE data, with both ONH cropping (C) and periphery cropping (D) policies applied.

Figure 3

Averaged saliency maps for a selected number of ONH crop experiments. Top row: glaucoma detection saliency, averaged over 2643 UZL test images; middle row: VCDR regression saliency averaged over 4765 UZL test images; bottom row: a complimentary UZL fundus image of a right eye for illustrative purposes. In the last column, we draw the sectors used by glaucoma experts to locate damage. Infero- and superotemporal (bottom- and top-left, respectively) sectors are the locations most commonly damaged by glaucoma.