Fully automated prediction of geographic atrophy growth using quantitative SD-OCT imaging biomarkers

Abstract

Purpose : To develop a predictive model based on quantitative characteristics of geographic atrophy (GA) to estimate future macular areas of GA growth.

Methods : Imaging features of GA quantifying its extent and location, as well as characteristics at each topographic location related to individual retinal layer thickness and reflectivity, drusen, and photoreceptor (PR) loss, were automatically extracted from 118 SD-OCT scans of 38 eyes in 29 patients collected over a period of 5 years with baseline GA areas of 0.18 to 19.5 mm². We developed and evaluated a model to predict the magnitude and location of GA growth using the quantitative features as predictors in three scenarios: (1) predicting GA growth at first follow-up visit in independent patients, (2) predicting GA growth at subsequent follow-up visits in independent patients, and (3) predicting GA growth at subsequent visits using prior outcomes from the same eye at the first follow-up visit.

Results : PR loss, reduced inner/outer segment (IS/OS) reflectivity and projected RPE layer were the most informative features for predicting future GA. The prediction scores for future GA presented an area under the receiver operating characteristic curve of 0.95, 0.97 and 0.96 in the three scenarios. The predicted GA regions in the three scenarios resulted in a Dice index (DI) mean of 0.81±0.12, 0.84±0.10 and 0.88±0.59, respectively, when compared to the observed ground truth. Considering only the regions without evidence of GA at baseline, predicted regions of future GA growth showed relatively high DIs of 0.69±0.18, 0.72±0.18 and 0.70±0.22, respectively. Predictions and actual values of GA expansion rate and future GA involvement in the central fovea were highly correlated.

Conclusions : Thinning and loss of reflectivity of the IS/OS seems to be the most significant early indicator of regions that are susceptible to GA expansion. Our predictive model has the potential ability to identify the regions where GA is likely to appear.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

(A) baseline with segmented GA (white). (B) 12 months from baseline. The GA region at baseline (white), observed GA (red), predicted GA (blue). (C) Comparison of prediction results vs observed GA. Correctly predicted (TP, red) and incorrectly predicted (FP, white) GA regions. Missed expansion regions in the predictions (FN, black). (D) Classification probability from evaluating the baseline scan.