Abstract
Purpose :
To develop a deep learning system that can, from OCT b-scans, capture changes in retinal layer thickness over time and predict the progression of dry to wet age-related macular degeneration (AMD).
Methods :
We used a dataset of 2032 Optical Coherence Tomography (OCT) images (496x768 resolution) collected from 188 patients clinically identified with early and intermediate AMD and followed over a period of time (mean 10.8 years, std 8.78). The region between retinal pigment epithelium (RPE) and outer plexiform layer (OPL) was first segmented by experts, and then its thickness was measured by averaging over the horizontal axis the number of pixels between the two layers. Next, a Convolutional-Neural Network (CNN) was trained to capture the thicknesses from the longitudinal OCT scans, formulating a thicknesses time series for each patient. At last, a long-short-term memory (LSTM) neural network model was trained to process the time series and predict the thickness change in the future.
Results :
Sequence segments from 138 patients (1310 OCT images), progressing from early stages to wet AMD (just before anti-VEGF treatment initiation), were used to train and test the CNN and LSTM model (following an 8:2 ratio). The rest 722 images were only used to train the CNN. For all the sequences in the test set, images obtained from t=0 to t=T-1 were used to evaluate the performance of the CNN model, while t=T was left for evaluating the LSTM model. The CNN model achieved a 9.81% mean absolute error (MAE) between the predicted thicknesses and the measurements extracted from segmentations. The LSTM model achieved 15.72% MAE in terms of predicting the RPE-OPL layer thickness at t=T, using the thickness time series (t=0 to T-1) extrapolated by the CNN model.
Conclusions :
Our learning-based approach could facilitate longitudinal prediction of retinal layer thickness changes by analyzing OCT scans.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.