Automated Measurement of RNFL Thickness from OCT Scans Using a Deep Learning Segmentation Free Approach

Abstract

Purpose : Analysis of the retinal nerve fiber layer (RNFL) on spectral domain optical coherence tomography (SD-OCT) is a vital part of the diagnostic workup and management of glaucoma. Deep learning (DL) models often perform automatic segmentation of the RNFL to provide diagnostic assistance for glaucoma. However, each instrument uses proprietary algorithms for their retinal layer segmentation. A segmentation free approach may facilitate harmonization of RNFL thickness across instruments. The purpose of this study is to assess whether a segmentation free approach can be used to train a DL model to automatically measure the RNFL thickness maps from OCT.

Methods : A set of 22,464 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 were gathered. Data were split on the patient level into training (571 patients, 969 eyes, 20,528 images) and testing (48 patients, 83 eyes, 1936 images) sets. Each circle scan had a corresponding RNFL thickness map consisting of 768 measurements in µm, with an average thickness of 83.0 +/- 21.0. µm. Three commonly used convolutional neural network (CNN) model architectures, ResNet50, Densnet121, and EfficientNetB0 were trained to predict the 768 thickness values from each corresponding OCT scan. Mean absolute error (MAE) was used as the loss function. Performance was measured using MAE, mean squared error (MSE), and root mean squared error (RMSE) averaged across all 768 thickness values.

Results : Results for automated RNFL thickness measurement are shown in figure 1. EfficientNetB0 achieved modestly better performance than ResNet50 and DenseNet121 with MAEs of 20.2, 21.4 and 20.0 um respectively on the testing set. Training and testing loss curves for the ResNet50 model are shown in figure 2. Upon initialization, the model begins with an average MAE of roughly 80 µm, and within 20 epochs, converges to an MAE of roughly 20 µm.

Conclusions : A DL model can automatically measure RNFL thickness from OCT with reasonable performance, without using segmentation labels. A segmentation free approach may be more feasible for training deep learning models, and merits further investigation.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Figure 1: Results on RNFL thickness measurement for the testing set.

 

Figure 2: Loss curve for ResNet50, measured in MAE (micrometers) for given number of epochs.