Artificial intelligence to differentiate smokers and non-smokers using fluorescence lifetime imaging ophthalmoscopy

Abstract

Purpose : Fluorescence lifetime imaging ophthalmoscopy (FLIO) has been shown to have the potential to detect metabolic changes in the fundus. We have already reported that smokers and non-smokers exhibit slightly but significantly different fluorescence lifetime (FLT) of ocular fundus, strongly suggesting that FLIO may be highly sensitive in detecting early retinal metabolic changes (Kreikenbohm et al, ARVO2022). However, the clinical application with this apparently normal FLIO image requires a high degree of discriminative ability. Therefore, we investigated the possibility of implementing an artificial intelligence (AI) approach to FLIO data.

Methods : The FLIO data were obtained from 28 smokers (more than 5 cigarettes/day, longer than 2 years) and 26 age-matched nonsmokers, healthy adults between 20 and 37 years of age with no systemic or retinal diseases. The mean FLT (τ_m) and fluorescence intensity data of all pixel positions of the central 30° were extracted as 256x256 matrix data. After these data were preprocessed using a standardized Early Treatment Diabetic Retinopathy Study (ETDRS) grid, which indicates a central, inner, and outer ring, the data were analyzed using support vector machines (SVM) with a non-linear radial bases function kernel. For evaluation, the results were averaged over 20 iterations of 5-fold cross validation and accuracy was corrected for data imbalance. In addition, OCT-Angiography data from the same subjects were also evaluated using the local fractal dimension to assess vessel density and with the SVM classification framework developed for FLIO data.

Results : The implemented AI approach was able to distinguish FLIO data (τ_m) between non-smokers and heavy smokers with about 80% accuracy. The highest performance was achieved for two feature sets: 1) the τ_m of inner ETDRS ring in the short spectral channel and outer ring in the long spectral channel (79.56%), and 2) the τ_m at temporal and superior of the inner ring in the short spectral channel and at superior of the outer ring in the long spectral channel (80.00%). Regarding OCT-Angiography data, no differences were shown between groups.

Conclusions : The AI approach showed high accuracy of distinguishing smokers and non-smokers in young healthy adults. AI-aided FLIO may greatly advance the diagnosis of early retinal metabolic change.

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