A Deep Learning Approach Accurately Estimates RNFL Thickness in OCT Scans Where Instrument-Native Segmentation Fails

Abstract

Purpose : Retinal nerve fiber layer (RNFL) analysis from optical coherence tomography (OCT) is crucial to the diagnosis and treatment of glaucoma. Established models use proprietary algorithms for retinal layer segmentation, which are susceptible to error. The purpose of this study is to evaluate the performance of 3 segmentation-free deep learning (DL) models, trained to estimate RNFL thickness from unsegmented OCT imaging in both cases where traditional segmentation was correct and incorrect.

Methods : 25,371 OCT Spectralis optic nerve head (ONH) circle scans from the Diagnostic Innovations in Glaucoma Study (DIGS) and African Descent and Glaucoma Evaluation Study (ADAGES) datasets from 981 patients (1567 eyes) were collected. This dataset was then divided into 3 groups by patient: a training (20,528 images) set, a test set composed of OCT images that required no manual segmentation correction (n=1936), and a test set that required manual segmentation (n=2907). These circle scans were each associated with 768 RNFL thickness values, used as ground truth. Three DL architectures, ResNet50, Densenet121, and EfficientNetB0, all trained on the common training set were used to predict RNFL thickness from unsegmented OCT circle scans. They were applied to both testing sets, with OCT images where segmentation had either succeeded (n=1936 scans) or failed (n=2907 scans). Performance was defined as mean absolute error (MAE) and root mean squared error (RMSE) compared to the instrument software RNFL thickness measurement.

Results : Performance did not significantly differ between the two test sets (MAEs 20.24, 21.39, 19.97 vs 19.11, 19.44, 19.8; RMSEs 26.77, 29.09, 26.56 vs 25.30, 26.55, 25.27). Stratification by race, sex, age, and glaucoma status did not demonstrate significant differences in performance within these testing sets.

Conclusions : A DL model can accurately determine RNFL thickness from OCT images without explicit segmentation across patient demographic and diagnostic groups, even in scans where existing segmentation approaches failed. This segmentation-free method enables accurate RNFL thickness estimation in the presence of segmentation errors, potentially reducing the impact of such errors and the need for manual correction.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1: MAEs and RMSEs for ResNet50, DenseNet121, and EfficientNetB0 for the test sets.

 

Table 1: Clinical and demographic characteristics of the data sets.