Automated Deep Learning-based Disease Feature Quantification on Color Fundus Photographs for Prediction of Late-stage Age-related Macular Degeneration

Abstract

Purpose : Intermediate age-related macular degeneration (AMD) is characterized by many soft drusen often accompanied by pigment changes; this stage is a prominent risk factor for development of late stages. Deep learning (DL) algorithms have thus far only paid attention to detection and quantification of drusen diagnosed on OCT. Since most screening programs use only color fundus photographs, and performance of prediction models for progression may benefit from inclusion of all intermediate biomarkers, we developed a DL model to automatically detect and quantify drusen, hyperpigmentation and retinal pigment epithelium (RPE) degeneration (hypopigmentation) on color fundus photographs (CFP).

Methods : As a pilot, we selected 380 participants aged 65+ years with high quality digital macula-centered CPFs from the 5th round of the population-based Rotterdam Study using a DL algorithm, of whom 43 developed late AMD during an average follow up time of 4.9 years (SD 0.8). Drusen, hyperpigmentation and RPE degeneration lesions were automatically detected and their size and total area were quantified within the Early Treatment Diabetic Retinopathy (ETDRS) grid with a DL algorithm. Area under the receiver operating characteristic curve (AUCs) were constructed for models predicting progression to late AMD at next follow-up.

Results : Our late AMD prediction model based on automated DL-based drusen quantification achieved an AUC of 0.89 (95% CI 0.84-0.95). The addition of quantification of hyperpigmentation and RPE degeneration to our model achieved an AUC of 0.92 (95% CI 0.87-0.98).

Conclusions : Automated disease feature quantification using DL algorithms appears to be effective for prediction of progression to late AMD. Inclusion of pigment changes improved a drusen-only prediction model. When applicable to images of all qualities, our model can help identify persons at risk of late AMD using only color fundus photography.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.