Abstract
Purpose :
Absence of gold-standard definition for progression hinders the development of supervised artificial intelligence algorithms for detecting glaucoma progression. In this research we develop a weakly-supervised time series deep learning (DL) algorithm to detect glaucoma progression in the absence of ground truth using a series of optical coherence tomography (OCT) images.
Methods :
The study included glaucomatous and healthy eyes with a minimum of 5 reliable OCT scans during follow-up. Two convolutional neural networks+long short-term memory (CNN-LSTM) models with parallel learning schemes were trained to classify whether OCT B-scans' sequences showed progression. Only healthy eyes in the dataset (10%) were labeled (positive labels) and glaucoma eyes were not labeled for progression. The first model used positive unlabeled (PU) learning, trained as one-class classifier to identify sequence of tests from glaucoma vs healthy eyes. The second was noise learning model developed to differentiate original follow-up sequence of OCT tests (pseudo-labeled for glaucoma progression) from a subset of OCTs presented under randomly scrambled order (pseudo-labeled for non-progression). The features of PU and noise learning models were combined at classification stage and jointly trained to identify eyes with true glaucoma progression while accounting for normal age-related loss learned from healthy eyes. The DL model’s outcomes were evaluated against ordinary least square (OLS) regression of global retinal nerve fiber layer (RNFL) thickness over time. Comparison was performed on separate test set of glaucoma eyes where all original sequences of tests were available. The methods were matched for specificity, and the hit-ratios for detecting progression were compared.
Results :
8,785 follow-up sequences of 5 observations of 3,253 eyes from 1,859 participants were included in the study. Mean follow-up period was of 3.1 ± 2.2 years. Eyes had average rate of RNFL thickness change of –0.72 ± 1.56 µm/year, as calculated using OLS regression. The model trained on PU identified 48% of the original test sequences as progressing versus only 19% for OLS, both matched at a specificity of 95% (P<0.001).
Conclusions :
A DL model was able to identify glaucoma progression using structural changes observed in sequences of OCT B-scans in the absence of ground truth when trained with weak supervision.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.