Deep learning automated detection of reticular pseudodrusen from fundus autofluorescence images and color fundus photographs in the Age-Related Eye Disease Study 2 (AREDS2)

Abstract

Purpose : To develop and evaluate deep learning (DL) models for detecting reticular pseudodrusen (RPD) using fundus autofluorescence (FAF) images. Second, using the novel technique of label transfer, to develop models for RPD detection using color fundus photographs (CFP). The latter is important because RPD are detected by human graders at much lower sensitivity on CFP than on FAF.

Methods : The AREDS2 dataset was used, comprising 11,535 FAF images of 2,450 participants with age-related macular degeneration (AMD), and 11,535 corresponding CFP. Ground truth labels were from reading center grading of the FAF images; these were transferred to corresponding CFP images. A FAF DL model was trained to detect RPD from FAF images. Using label transfer, a CFP DL model was trained to detect RPD from CFP. The main outcome measure was area under curve (AUC). Additionally, model performance was compared with grading of a subset of FAF and CFP images by four retinal specialists in a clinical setting.

Results : The FAF model had AUC 0.939 (95% CI 0.927-0.950), accuracy 0.899 (0.887-0.911), F1-score 0.783 (0.755-0.809), and kappa 0.718 (0.685-0.751) (Figure A). The CFP model had values of 0.832 (0.812-0.851), 0.809 (0.793-0.825), 0.593 (0.557-0.627), and 0.470 (0.426-0.511), respectively (Figure B). Performance measures for the FAF model were generally superior to those of the four retinal specialist graders; they exceeded the mean and matched the highest scores of the four specialists (Table). Performance measures for the CFP model were substantially superior to those of the four retinal specialist graders (Table).

Conclusions : DL-enabled automated detection of RPD presence can be achieved at high accuracy from FAF images, with equal or superior performance to that of retinal specialists. DL detection from CFP images also has relatively high accuracy, with substantially superior performance to retinal specialists. DL models can assist, and even augment, the detection of this clinically important, AMD-associated lesion.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure. ROC curves for RPD detection by deep learning models from (A) FAF images and (B) corresponding CFP images. Human retinal specialist performance is shown (green dots).

 

Table. Performance of the deep learning models and four retinal specialists for RPD detection. For each metric, the highest result is highlighted in bold.