Artificial Intelligence in Fluorescence Lifetime Imaging Ophthalmoscopy (FLIO) Data Analysis-Toward Retinal Metabolic Diagnostics

Abstract

The purpose of this study was to investigate the possibility of implementing an artificial intelligence (AI) approach for the analysis of fluorescence lifetime imaging ophthalmoscopy (FLIO) data even with small data. FLIO data, including the fluorescence intensity and mean fluorescence lifetime (τm) of two spectral channels, as well as OCT-A data from 26 non-smokers and 28 smokers without systemic and ocular diseases were used. The analysis was performed with support vector machines (SVMs), a well-known AI method for small datasets, and compared with the results of convolutional neural networks (CNNs) and autoencoder networks. The SVM was the only tested AI method, which was able to distinguish τ_m between non-smokers and heavy smokers. The accuracy was about 80%. OCT-A data did not show significant differences. The feasibility and usefulness of the AI in analyzing FLIO and OCT-A data without any apparent retinal diseases were demonstrated. Although further studies with larger datasets are necessary to validate the results, the results greatly suggest that AI could be useful in analyzing FLIO-data even from healthy subjects without retinal disease and even with small datasets. AI-assisted FLIO is expected to greatly advance early retinal diagnosis.

Keywords: artificial intelligence; fluorescence lifetime imaging ophthalmoscopy; retinal metabolism; small data; smoking; support vector machine.

Figure 1

Representative FLIO data from a non-smoker (left) and a smoker (right). FLIO Data: Pseudo-colored images of mean fluorescence lifetime (τ_m) (top) and intensity images (bottom). All show typical findings. The pseudo colors reproduce τ_m in a range of 190 to 350 picoseconds (ps; see color legend).

Figure 2

En face vascular images from the 15 default segmentations in OCT-A: Full: all layers, Vitreoretinal interface, Retina, SVC: superficial vascular complex, NFLVP: nerve fiber layer vascular plexus, SVP: superficial vascular plexus, DVC: deep vascular complex, ICP: intermediate capillary plexus, DCP: deep capillary plexus, Avascular complex, CC: choriocapillaris, choroid, HL: Haller’s layer, ILMtoBM40: internal limiting membrane to Bruch membrane, and SL: Sattler’s layer.

Figure 3

Processing of FLIO data for AI-based analysis. (A) For every grid section, the mean fluorescence lifetime (τ_m) and fluorescence intensity value were computed for the extraction of features according to the grid of the early treatment diabetic retinopathy study (ETDRS), with the rings of the central area (C), inner ring (IR), and outer ring (OR), and the further division into the 4 subareas: nasal (N), superior (S), temporal (T), and inferior (I). (B) Schematic of the flow of the AI-based analysis of FLIO data. The workflow first obtains the data as 256 × 256 matrices for fluorescence intensity and fluorescence lifetime measurements for both the SSC and LSC, then obtains the means over the sectors of the ETDRS grid (sectorization), combines the means of all four matrices into one vector, and finally learns a classification with an SVM based on all data points from the dataset.

Figure 4

Schematics of (A) a general artificial neural network (ANN) and (B) an autoencoder network. The input layer matches the size of the input data to the neural network. ANNs can then have several intermediate layers with autoencoders typically having a central layer that is small compared to the other layers (a bottleneck). For classification tasks, the output layer matches the number of classes. For autoencoders, the output layer matches the size of the input layer.

Figure 5

Schematic of the flow of the AI-based analysis of OCT-A. The workflow starts with the OCT-A image data. For some experiments, vessel density maps were calculated using local fractal dimensions. The data were then further processed using the means over the sectors of the ETDRS grid (sectorization) and learning classifications using SVMs or calculating histograms over the measurements or using neural networks to obtain feature vectors that were then compared between smokers and non-smokers using t-SNEs. All four possible combinations were executed for each analyzed sample.

Figure 6

t-distributed stochastic neighbor embedding (t-SNE) plots of different image encodings with the means per group (+) and standard deviations (ovals) of OCT-A data. (A) Histogram encoding; (B) Neural Network encoding; (C) Sectorization encoding; (D) Sectorization encoding on a density map. The axes comprise the two dimensions (Dim 1, Dim 2) of the embedding space of the t-SNE and are therefore dimensionless and only describe two directions, where similar points from the origin space are clustered and dissimilar points are far from each other.