Retinal pulsatile shift enhancement with blind source separation

Abstract

Purpose : Spontaneous venous pulsation (SVP) and optic nerve head (ONH) are a dominant hemodynamical phenomena in retina measured as reflectance intensity changes in semi-automatically segmented regions of interest (ROIs) in a video-ophthalmoscopic recording. We are innovatively implementing a blind source separation (BSS) of pulsatile components, which automatically extracts SVPs and other pulsations in ONH.

Methods : Video sequences (480x640px, 25fps, time=10s) of mydriased left eyes were recorded in 51 participants (glaucoma patients and healthy controls). Recorded sequences were co-registered and detrended. Two different BSSs and one common averaging (AVG) approach were utilized on ONH data. The 1^st BSS implements spatial Principle Component Analysis (sPCA) and decomposes the data at 5 PCs. The 2^nd BSS uses sPCA for dimensionality reduction and follows with spatial Independent Component Analysis (sICA) estimating 5 ICs. The 3^rd approach estimates pulsatile averages from two different ROIs (i.e. vessels in ONH, rest of the ONH). Resulting curves were low pass filtered (f<9.4Hz). Pulsatile fluctuations were extracted and classified by visual inspection. Pearson correlation coefficients (r) evaluated relations between pulsatile pairs. Time shifts (s) related to cardiac cycle period were calculated over pulsatile pairs. Differences in both r and s parameter group distributions were evaluated over three used approaches by Wilcoxon test.

Results : BSSs spatially separated central SVPs and ONH temporal parts with anti-correlated vessels. Evaluated r for SVP-Disc pairs showed lower correlation for BSS techniques (median ± standard deviation – AVG 0.9±0.1, sPCA 0.6±0.3, sICA 0.6±0.3). The s ratio between ONH and SVP demonstrated higher degree of dephasing for pulsations derived with BSSs (AVG 0.0±2.1%, sPCA -7.7±9.6%, sICA -7.9±11.8%). Evaluated parameters (r, s) of both BSSs significantly differed from values derived with AVG (sPCA p_r=4e^-14, p_s=3e^-7; sICA p_r=5e^-13, p_s=8e^-8).

Conclusions : We have innovatively utilized sPCA and sICA for blind identification of different retinal pulsations. BSS techniques identified minimally 2 pulsatile patterns including SVP and other ONH pulsations in each subject. AVG demonstrated a presence of SVP and highly correlated ONH pulsations. BSSs showed to be superior to the AVG as they decrease the degrees of correlations and phase coherences over different spatially orthogonal or independent patterns.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.