Abstract
Purpose :
To demonstrate that hyperspectral retinal imaging can be used in combination with deep learning to provide accurate estimates of retinal nerve fibre layer (RNFL) thickness.
Methods :
We recruited 47 people with glaucoma and 47 healthy controls and imaged them with a hyperspectral retinal camera (Optina Diagnotics, Metabolic Hyperspectral Retinal Camera; wavelength range 450-900 nm) and with an OCT (Heidelberg Spectralis; macular scan). Machine-learning was used to co-register OCT and hyperspectral images. RNFL thickness at each retinal location (per pixel) was extracted from the segmented OCT map for each participant and this was used as the ground truth for a deep learning algorithm trained with hyperspectral image data. Spectral data (images acquired at 91 wavelengths per patient) were the input for a spectral convolutional neural network comprised of convolutional inception layers, one fully connected layer in the hidden layers and a regression layer optimizing for the output layer root mean square error (RMSE). Restricting the input to spectral data alone limited the number of tuneable parameters (<16k), making model training more efficient and less prone to overfitting. The dataset was randomly split into 80 images for training (>400K individual spectra) and 14 images for testing (>70K spectra). Learning was tuned to minimize the RMSE without overfitting the training set.
Results :
RMSE was 13μm for the training set and 15μm for test set. A high and significant pixel-wise correlation was found between OCT and hyperspectral RNFL thicknesses (r=0.8, p<1e-20 for training and 0.74, p<1e-20 for testing). The limit of agreement (Bland-Altman plots) for mean RNFL measures was 12μm. No significant bias of mean thickness measurements was found (r = 0.06, p =0.56). These findings are in keeping with the inter-device limits of agreement reported for OCT RNFL measurements.
Conclusions :
Hyperspectral imaging macular RNFL thickness estimates were similar to OCT measurements in the cohort studied. Further studies are required to determine the robustness of this method in a larger cohort, including individuals with macular disease.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.