An end-to-end deep learning method for predicting retinal layer thickness from Optical Coherence Tomography (OCT) images in patients with clinical progression of Age-related Macular Degeneration (AMD)

Abstract

Purpose : To develop a fully-automated, end-to-end system that can capture and predict changes in retinal layer thickness from OCT scans.

Methods : We used 1308 OCT images (496 x 768 resolution) collected from 118 patients with AMD that progressed over time. Each patient dataset contained a consecutive sequence of OCT images obtained from multiple follow-up visits over at least two years (mean follow-up time 11.08 years ± 8.53). The region between the retinal pigment epithelium (RPE) and outer plexiform layer (OPL) was initially segmented by experts, and its thickness was presented by averaging the number of pixels between the two layers over the horizontal axis. Then, a Convolutional-Neural Network (CNN) was trained to capture the thicknesses from the longitudinal OCT scans, formulating thicknesses time-series for each patient. At last, a long-short term memory (LSTM) model was trained to process the time series and predict the thickness change in the future.

Results : OCT sequences in the dataset are split by 8:2 ratio to constitute training and testing dataset. In the test set, images obtained from t=0 to t=T-1 were used to evaluate the performance of the CNN, while t=T was left for evaluating the LSTM. The CNN achieved 13.0% mean absolute error (MAE) between the predicted thicknesses and the measurements extracted from segmentations. The LSTM achieved 21.46% MAE in predicting the RPE-OPL layer thickness at t=T, using the thickness time-series (t=0 toT-1) extrapolated by the CNN.

Conclusions : Our learning-based approach can facilitate predictions of retinal layer thickness changes directly from the OCT scans that could be used to predict AMD progression or response to treatment.

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

 

Training flow: (i) RPE-OPL layer segmentations was generated from the sequence of OCT images, from which the layer thickness was quantified. (ii) The CNN model was trained to capture the RPE-OPL layer thickness from the images in the sequence (t=0 to T-1). (iii) An LSTM model was used to predict the thickness at t=T.

 

Visualization of the layer-thickness time-series and predictions for all the patients. The RPE-OPL thicknesses extracted from the segmentations are shown in black or blue. The thickness captured by the CNN model shown in green. The predicted layer thickness for t=T, by the LSTM model, are shown in orange.