Non-invasive detection of retinal ischaemia using hyperspectral imaging

Abstract

Purpose : To determine whether hyperspectral retinal imaging can be used in combination with deep learning to detect retinal ischaemia.

Methods : We performed hyperspectral retinal imaging (Optina Diagnostics, Metabolic Hyperspectral Retinal Camera; wavelength range 450-900 nm) on 23 people with retinal ischaemia due to diabetic retinopathy or retinal vein occlusion and 32 people without ischaemia, prior to fundus fluorescein angiography (FFA) imaging. Avascular retinal areas on angiogram images were annotated by a retinal specialist using custom cross-modality annotation software. Machine learning was used to co-register annotated FFA and hyperspectral images. Data were randomly split into 45 images for training (>3.1M spectra) and 10 images for testing (>0.7M spectra). Spectral data (images acquired at 91 wavelengths per patient) were the input for a convolutional neural network comprised of a convolutional inception layer, one fully connected layer in the hidden layers and a classification layer. The network was trained and optimized for classification accuracy (ischaemic vs non-ischaemic pixels), without overfitting the training set. Restricting the input to spectral data alone limited the number of tuneable parameters (<16k), making model training more efficient and less prone to overfitting.

Results : Classification accuracy (per pixel) was 90.8% for the training set and 88.2% for the test set. The area under the receiver operating characteristic curve was 0.93 for training and 0.86 for testing (Figure 1), indicative of close correspondence with FFA for the detection of retinal ischaemia.

Conclusions : This study has demonstrated that retinal hyperspectral imaging can be used to identify areas of retinal ischaemia with close correspondence to FFA. Larger studies are required to validate these preliminary findings.

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