Abstract
Purpose :
To develop a novel deep learning (DL) algorithm to predict SAP progression from longitudinal optical coherence tomography (OCT) scans.
Methods :
The study included glaucomatous eyes with ≥ 5 SAP tests and ≥ 5 OCT scans. A convolution neural network-long short-term memory (CNN-LSTM) DL model was trained to classify whether OCT B-scan sequences showed glaucomatous progression, using events of “likely progression” defined by the SAP-guided progression analysis (GPA) algorithm as reference. The training process involved three steps to increase the model performance to predict functional progression from structural data and reduce bias on the unbalanced data: 1) a binary classification CNN-LSTM model to predict GPA progression using original OCT sequences; 2) training on a subset of adversarially augmented, randomly shuffled sequences to create “hard negatives” for progression; and 3) applying self-supervised contrastive learning between original and shuffled datasets to discern “true” representations of change over time. The DL model’s outcomes were evaluated against progression determined by ordinary least squares (OLS) regression of retinal nerve fiber layer (RNFL) thickness over time by comparing sensitivities at matched 95% specificity.
Results :
649 OCT sequences from 614 eyes of 424 subjects were included in the study. The mean follow-up time was 7.0±2.9 years, and baseline SAP mean deviation was -4.0±5.2 dB. 9% or 56 of the original sequences (45 eyes) were classified as progressors based on the GPA. The DL model had an AUC score of 0.894 (95%CI; 0.825–0.963), with a sensitivity of 0.500 (95%CI; 0.492–0.508) at matched 95% specificity for the GPA criteria. In comparison, OLS regression for RNFL over time had a much lower sensitivity of 0.071 (95%CI; 0.067–0.076; P<0.001) at specificities matched at 95%.
Conclusions :
A DL model was able to better predict functional loss from structural tests than OLS regression by analyzing structural changes in sequential OCT B-scans.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.