Retinal Nerve Fiber Layer Thickness Prediction from Longitudinal OCT Imaging using Deep Learning

Abstract

Purpose : To employ a custom deep learning model (one dimensional convolutional neural network, 1D-CNN) to predict future retinal nerve fiber layer (RNFL) thickness based on prior optical coherence tomography (OCT) scans.

Methods : This study utilized optic nerve head (ONH) OCT circle scans (Spectralis, Heidelberg Engineering, Heidelberg, Germany) collected longitudinally from participants in the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES). From the longitudinal data, input-target series composed of three OCT scans with a gap of 6 – 18 months between each scan were used to train and evaluate models. RNFL thickness measurements from the first two scans of each series were used as model input to predict RNFL thickness in the third scan (Figure 1). Data was split by participant into training / validation (691 input-target series, 357 eyes, 205 patients) and testing (126 input-target series, 80 eyes, 46 patients) sets. The custom 1D-CNN model included was compared to linear regression (LR), support vector regression (SVR), and gradient boosting regression (GBR) models.

Results : The 1D-CNN model significantly outperformed GBR, SVR, and LR in predicting RNFL thickness. (Table 1). For the global and 6 sectoral average thicknesses, it achieved a mean absolute error (MAE) ranging between 2.03μm (global) to 4.64μm (temporal superior), mean relative error (MRE) ranging between 3.27% (global) and 5.44% (TI), and an R-Squared between 0.94 (nasal) an 0.97 (global). The model's predictions were robust and did not show statistically significant variation across different gender, race, and disease severity. In predicting all 768 RNFL thickness measurements from the circle scans, the 1D-CNN and GBR models achieved the highest accuracy (MAE = 6.69μm and 5.21μm, respectively).

Conclusions : This study demonstrates the effectiveness of a custom 1D-CNN model in predicting future RNFL thickness from prior OCT imaging. Accuracy is maintained across different patient demographics and disease severities. The high accuracy in predicting and mapping RNFL thickness changes shows potential for predicting glaucoma progression to improve glaucoma management.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

Figure 1: Our method for predicting future RNFL thickness from previous OCT imaging.

 

Table 1: Comparison of models to predict global and sectoral RNFL thickness averages.