Retrieving retinal autofluorescence from retinal hyperspectral images using deep neural network analysis.

Abstract

Background: Hyperspectral imaging (HSI) of the retina has been used to detect novel disease biomarkers. It is not known if hyperspectral imaging can serve as a surrogate of fundus autofluorescence (FAF) imaging. Conventional FAF imaging requires prolonged exposure to intense short wavelength light which limits patient acceptability and clinical utility. We hypothesize that the retinal spectral reflectance measured with HSI can be used to construct an accurate simulation of FAF. Aim(s): To devise a method for estimating retinal autofluorescence using HSI. Method(s): Hyperspectral and FAF images of 64 patients with age-related macular degeneration were co-registered using a machine learning software tool such that corresponding areas could be viewed and annotated simultaneously. FAF images were manually annotated to mark areas of hyperfluorescence, hypofluorescence and normal fluorescence. A multiclass 1-dimensional convolutional neural network was then trained using hyperspectral images of 54 participants, using the FAF annotations as the ground truth for fluorescence. Images of 10 participants were kept aside for testing. The trained deep learning network was then used to construct maps of retinal autofluorescence from the retinal hyperspectral images. Result(s): The analysis is in progress and will be complete in the coming weeks. Pilot studies indicate close correspondence between FAF and hyperspectral fluorescence maps. Conclusion(s): This project sets the foundation for further research into retinal fluorescence analysis with HSI, potentially enabling quantitative measurement of fluorescence and in vivo sub-typing of retinal fluorophores.