Linking Function and Structure: Prediction of Retinal Sensitivity in AMD from OCT using Deep Learning

Abstract

Purpose : Microperimetry examines retinal sensitivity at specific retinal locations and is the most comprehensive test of visual function at the macular level. However, it is also time consuming and has limitations in reproducibility. To establish reliable morphologic surrogate endpoints for visual function testing, we developed and evaluated a deep learning model to predict retinal sensitivity maps (function) from OCT volumes (structure). The consolidated use of microperimetry and deep learning offers the opportunity to detect novel biomarker candidates for retinal function, to evaluate to which extent retinal function can be explained by structure shown in OCT, and to investigate which features of the retina affect its sensitivity. Methods : As a representative dataset, we used 463 pairs of SD-OCT volume scans (512x128x1024 voxels, Cirrus, Zeiss) and corresponding microperimetry examinations (Nidek MP1), obtained in 174 patients with a healthy retina, early or intermediate AMD, CNV or GA. First, registration of each OCT-microperimetry pair was performed. Then, a deep learning model was trained to predict an en-face 2D retinal sensitivity map for a 3D OCT volume, in which a retinal sensitivity value is assigned to each A-scan location. Results : We compare three different models, using A-scans (DL-1), B-scans (DL-2) or 3D patches (DL-3) as input for prediction. We achieved an overall Mean Absolute Error (MAE) for point-wise sensitivity (PWS) of 3.07 dB (DL-1), 2.87 dB (DL-2) and 2.67 dB (DL-3) on the test set. To investigate which parts of the OCT were important for retinal sensitivity prediction, explanatory techniques to understand the model decisions can be applied. Conclusions : We propose a deep learning methodology to predict retinal sensitivity from OCT volumes. Using a model based on three-dimensional input improves the prediction performance, indicating the importance of context information. Subsequent qualitative evaluation using saliency maps enable an in in-depth evaluation. The results of this study are a promising step in exploring the linkage between image-based information and function, in particular in the context of novel biomarker candidate detection.