Deep-learning Embedding of Fluorescence Lifetime Imaging Ophthalmoscopy (FLIO) for Macular Telangiectasia Type 2 and Age-Related Macular Degeneration

Abstract

Purpose : Fluorescence lifetime imaging ophthalmoscopy (FLIO) allows for unprecedented temporal resolution of the macular fluorescence signal. Current visualization techniques for FLIO rely on hand-adjusted parameters that may not be reproducible. Here we fully-automate the visualization of FLIO data by embedding in a lower dimensional space, merging 1-D autoencoder and classification deep learning networks.

Methods : FLIO images were collected from healthy (n=91), MacTel (n=128) and AMD (n=30) eyes for the short (f1) and long (f2) spectral channels (each with 1024 time bins). The images were patient-split 80/20 into training and validation sets. Labels for vessels were segmented using a threshold on the mean across time. A parafoveal window of 64x64 pixels was extracted from each train subject and each pixel time course (f1 and f2) was truncated to 512 time points and labeled as either: Healthy, MacTel, AMD, or vessel. The f1 and f2 time courses were embedded into two dimensions with parallel autoencoders linked through a 4-class classification head (see Figure 1A). The whole validation images (256x256) were embedded into four dimensions (two each for f1 and f2) and visualized.

Results : We achieved a validation accuracy of 80.0% and correlation coefficients of 0.911 and 0.957 between the true and decoded time courses for f1 and f2, respectively (Fig 1B). The examples of the decoded time courses are more smooth but preserve the shape of the true FLIO time courses (Fig 1C). The example MacTel embeddings show changes in the perifoveal area while the example AMD embeddings show a more widespread alteration throughout the macula (Fig 2).

Conclusions : Utilizing autoencoders linked with a classification head, we created an embedding that is sufficient to recover the original signal (autoencoder) and is designed to be disease specific (classification). This fully-automated approach may enable FLIO to be used as a biomarker in future clinical studies.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Fig 1: A) Model architecture B) Training curves, classification accuracy (top) and correlation coefficient between true and decoded time (bottom) C) Example decoded time courses for one pixel in a Healthy, MacTel, and AMD eye from the validation set.

 

Fig 2: Example embeddings for a Healthy, MacTel and AMD eye from the validation set. The right-most column provides the pixel-level 4-class classification results.