Spatial-spectral characterization and mapping of labeled drusenoid deposits in non-neovascular age-related macular degeneration

Abstract

Purpose : Although the biogenesis of drusenoid deposits is not yet fully understood, it is accepted that they contribute to the functional loss associated with Age-related Macular Degeneration (AMD). This exploratory study is aimed at identifying spatial-spectral features characteristic of soft drusen, hard drusen and reticular pseudodrusen in non-neovascular AMD. We hypothesize that spatial-spectral features extracted from hyperspectral retinal images should permit discrimination between the different types of drusenoid deposits.

Methods : Hyperspectral retinal images captured between 450-905nm in steps of 5 nm, as well as corresponding Optical Coherence Tomography (OCT) scans, were collected for 152 eyes from 100 participants aged 50 and above, with non-neovascular AMD. Using the OCT scans as ground truth, 148 eyes from 97 participants were annotated by eye specialists for their drusenoid deposits. Images were then evaluated using local spectral normalization to identify characteristic spatial-spectral signatures, followed by the exploration of two classification approaches based on these signatures.

Results : Extraction of the mean spectrum of annotated drusenoid deposits after relative background normalization yielded a specific spatial-spectral signature for each drusenoid deposit type (Fig 1A). The annotated drusenoid deposits (Fig 1B) were classified based on their spatial-spectral signature using the Spectral Angle Mapping (SAM) method (Fig 1C). Classification of the annotated regions as drusenoid deposits or healthy retina based on this SAM method resulted in a sensitivity and specificity of 97% and 70%, respectively. Use of a Random Forest classifier provided even better classification, with sensitivity and specificity of 88% and 99%, respectively. Pixel per pixel classification using the SAM method further provided plausible maps of the drusenoid deposits and healthy retina even outside annotated areas (Fig 2B) when compared to the annotated image (Fig 2A).

Conclusions : A spatial-spectral signature was identified for each drusenoid deposit type which enabled promising classification performances. These data suggest that automatic identification of drusenoid deposits based on phenotypic features, rather than anatomical information, could become a reality, and improve patient risk stratification with further research.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.