Deep Survival Prediction of Progression from Intermediate to Atrophic AMD from Longitudinal Retinal OCT

Abstract

Purpose : In patients with intermediate age-related macular degeneration (iAMD), the risk of progression to
advanced stages is highly heterogeneous and the prognostic imaging biomarkers remain unclear. Building optical coherence tomography (OCT)-based deep learning prognostic models is challenging due to large dimensionality of OCT scans and limited size of longitudinal datasets. We train a deep survival model, to estimate from OCT an individual risk of conversion to macular atrophy.

Methods : The deep learning model was trained on a development dataset consisting of 267 eyes (1459 OCTs) with iAMD imaged every 6 months with Spectralis OCT (512x49x496 voxels). The model learns a probability of a conversion to macular atrophy, with a prognosis for every 6 months up to 36 months in the future. The model is trained on raw OCTs with Logistic Hazard (LH) loss, combining the advantages of survival models and deep learning. We compared our model with a standard deep learning classification approach trained with a Binary Cross Entropy (BCE), as well as with a traditional survival model with Cox Proportional Hazards (CPH) trained on a set of clinically relevant quantitative OCT biomarkers. The models were evaluated with a five-fold cross-validation. In addition, an independent longitudinal OCT dataset with 240 eyes (240 OCTs) was used as an external test set. The model performance was measured with Concordance Index (CCI).

Results : The prognostic deep survival models achieved a CCI of 0.79+/- 0.03 on the cross-validation, and 0.74 on the external test set. The resulting population risk estimates in the form of Kaplan-Meier curves for the converter and censored eyes (conversion beyond 36 months) on the development set are shown in Fig. 1. The BCE model achieved CCI of 0.78 +/- 0.06 on the cross-validation, and 0.73 on the external test set. Finally, the CoxPh model obtained CCI of 0.75 +/- 0.1 on the cross-validation, and 0.73 on the external test set.

Conclusions : Deep survival models allow building risk estimators for progression from iAMD to macular atrophy. They successfully combine the advantages of deep learning that can learn from raw OCT images with survival modelling paradigm that accounts for censoring and predictions for multiple time intervals.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.