Photoplethysmographic analysis of retinal videodata based on the Fourier domain approach

Abstract

Assessment of retinal blood flow inside the optic nerve head (ONH) and the peripapillary area is an important task in retinal imaging. For this purpose, an experimental binocular video ophthalmoscope that acquires precisely synchronized video sequences of the optic nerve head and peripapillary area from both eyes has been previously developed. It enables to compare specific characteristics of both eyes and efficiently detect the eye asymmetry. In this paper, we describe a novel methodology for the analysis of acquired video data using a photoplethysmographic approach. We describe and calculate the pulsatile attenuation amplitude (PAA) spatial map, which quantifies the maximum relative change of blood volume during a cardiac cycle using a frequency domain approach. We also describe in detail the origin of PAA maps from the fundamental (the first) and the second harmonic component of the pulsatile signal, and we compare the results obtained by time-based and frequency-based approaches. In several cases, we show the advantages and possibilities of this device and the appropriate image analysis approach - fast measurement and comparison of blood flow characteristics of both eyes at a glance, the robustness of this approach, and the possibility of easy detection of asymmetry.

Fig. 1.

The principle of the presented approach shows the main steps of the data processing – the acquired time stack is processed frame-by-frame [16] to align the sequence; distorted frames are detected [17] (providing indexes of frame numbers); the temporal signals for each x,y position are Fourier transformed to the spectral domain (missing values due to distorted frames are temporally interpolated and spatial filter is applied on each frame before the Fourier transform).

Fig. 2.

Relation between detected light intensity (top curve) and blood volume (bottom curve).

Fig. 3.

A) An example of 10 plethysmographic signals of 10 neighbouring pixels directly extracted from a recorded video sequence from a specific location on the retina. B) corresponding averaged magnitude spectrum with visible 1^st and 2^nd harmonic components. The strong low frequency components below 20 bpm represent the fluctuation of the mean value. C) An example of a pulsation signal extracted from a small region inside ONH is shown as yellow color (with interpolated missing values, before trend correction) and the original signal with artifacts is shown as a black curve. The red curve shows the low frequency trend signal I_avg(n) (see Eq. (4)). Here, the light intensity is shown. For the blood volume, the signal has to be inverted (see Fig. 2).

Fig. 4.

Results of a parallel measurement (right and left eye) for Subject 1 (without eye disease). A: color fundus image, B: average images of registered video sequences, C: calculated PAA₁ map computed according to [Eq. (7)] for each pixel, D: calculated PAA₁ map color coded. The yellow circle shows the position of the ONH for better orientation and comparison between different imaging modalities. The numbers in C correspond to the type of different regions, which can be typically found in all PAA maps (see text for more details).

Fig. 5.

Results of a parallel measurement (right and left eye) for Subject 2 with OHT on the right eye and perimetric glaucoma on the left eye. A: color fundus image, B: average images of registered video sequences, C: calculated PAA₁ map computed according to [Eq. (7)] for each pixel, D: calculated PAA₁ map color coded. The yellow circle shows the position of the ONH for better orientation and comparison between different imaging modalities.

Fig. 6.

Examples of parallel measurements for Subjects 1 and 2 (from top to bottom): fundus images; the 1^st harmonic component PAA₁(x,y); the 2^nd harmonic component PAA₂(x,y) (please note the halved magnitude of the colormap); sum of the 1^st and 2^nd with white inpainted blood vessels and ONH borders. The white and black arrows show the position of some main veins and arteries, respectively.

Fig. 7.

Examples of parallel measurements for two subjects 3 and 4 (from top to bottom): fundus images; the 1^st harmonic component PAA₁(x,y); the 2^nd harmonic component PAA₂(x,y) (please note the halved magnitude of the colormap); sum of the 1^st and 2^nd with white inpainted blood vessels and ONH borders. The white and black arrows show the position of some main veins and arteries, respectively. The blue arrows for the subject on the left indicate a moon-shaped area of peripapillary atrophy (zone beta) around ONH.

Fig. 8.

Examples of time-based and frequency-based methods. Five PAA spatial maps were computed by time-based method (second row) and frequency-based method (third row); the summation PAA₁+PAA₂ is presented. The first row shows the averaged image from the acquired sequence.

Fig. 9.

Bland-Altman plot for frequency-based and time-based methods for PAA estimation. A dependence of PAA differences on PAA average values is noticeable. The higher PAA difference occurs for higher PAA values. The total average PAA difference between the frequency-based and time-based approach is equal to 0.61%A. The units %A denotes percent attenuation.

Fig. 10.

An example of efficient fusion of RGB fundus image with PAA map for Subject 1 on the left side and Subject 2 on the right. The red channel was replaced by the PAA₁₊₂ map. The red color inside and around the darker blood vessels represents the vein pulsation and the red color near the lighter blood vessels depicts the artery bending. The CLAHE method was applied on the green and blue components of the fundus image. The black arrows point to bending blood vessels (arteries in these examples) and the white arrows point to pulsating blood vessels (veins in these examples). Visualization 2 also shows bending and pulsating vessels with the same color-coded arrows.