Abstract
Purpose :
the development of a deep learning algorithm to detect OCT-determined geographic atrophy (GA) and to predict the progression from intermediate age-related macular degeneration (AMD) to GA in the short-term on spectral domain optical coherence tomography (SD-OCT) scans.
Methods :
We analyzed SD-OCT imaging data from 316 patients with iAMD enrolled in the longitudinal, observational Age-Related Eye Disease Study 2 (AREDS2) Ancillary Spectral-Domain Optical Coherence Tomography (A2A) Study with adequate imaging for repeated measures. We designed an end-to-end, position-aware volumetric image classification model motivated by the observation that scan locations within a SD-OCT have different levels of predictive power for GA progression prediction. The model was developed with a proactive pseudo-intervention learning strategy to jointly allow for robust learning with small sample size and interpretation of predictions via saliency (attention) maps. To improve model performance given the small sample-size of the available dataset, we employed a second publicly available dataset consisting of 108,312 individual OCT B-scans from 4,686 individuals with various retinal diagnoses. The main outcome measure was the predictive performance for GA detection and progression to GA in 1 year, examined based on the area under the curve (AUC) of the receiver operating characteristic.
Results :
The proposed deep learning model (multi-scan position-aware model) outperforms simpler image-based models and a model based on qualitative and quantitative features of SD-OCT provided by human experts. The AUC for GA prediction was 0.945 for current year, and 0.937 for the following year. The addition of expert annotated features only improved the AUC for GA prediction by an AUC of 0.008. The method automatically identified and highlighted specific structural features on OCT most predictive of GA, essentially opening the “black box” of this AI algorithm.
Conclusions :
Our deep-learning method had excellent performance characteristics for detecting GA and for predicting progression from intermediate AMD to GA in the short-term (one year). Further validation in additional, independent datasets will be needed to determine the utility of this algorithm for prediction of vision-threatening nonexudative AMD and potential modifications for outcome prediction in other retinal diseases.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.