The application of modified camera systems that image the retinal vessel at two or more wavelengths has been reported in recent literature. Early in vivo measurements of retinal vessel oxygen saturation have been reported using a dual-wavelength measurement approach.
7–9 However, the use of multispectral imaging is now gaining pace, with recent work showing success in image acquisition and oxygen saturation acquisition.
6,10–13 Indeed, models based on only two or three wavelengths may be too simplistic to reliably and accurately provide retinal oximetry measurements, as the light reflectance measurements of the blood vessels are complex and influenced by a myriad of factors (including the diameter of the vessel, retinal pigmentation, lens yellowing, light scattering and autofluorescence occurring in the vitreous and cornea, and illumination uniformity). Models based on spectral-rich datasets are expected to be much more powerful in that respect.
14,15 It was suggested that increasing the number of discrete wavelengths at which the ocular fundus reflectance is measured could increase, by the square root of the considered wavelength number, the precision of the blood oximetry evaluation.
16 To our knowledge, however, no study compared directly the experimental results obtained with a two- or three-wavelength method with a method based on hyperspectral datasets.
The study demonstrated that hyperspectral imaging can be achieved using a modified nonflash camera equipped with a high-definition CCD camera and TLS. The data reveal promising within-session repeatability of the current HRC system for repeated hyperspectral imaging within session (COR 0.02–0.11 OD units for arterioles and 0.03–0.14 OD units for venules). This compares well to current commercially available two-wavelength instruments that have quoted ICC values between 0.91 and 0.94 for retinal branch arteries and between 0.84 and 0.88 for retinal branch veins.
17
Other reproducibility data report an SD of 3.7% in arteries and 5.3% in veins,
18 whereas recent data for the same system showed an SD for repeated measurements of 1.0% and 1.4% in retinal arteries and veins respectively.
19 The former values fall very similar to those reported in our study and are as expected with a new prototype system.
Also consistent with others,
17 venous values exhibit more deviation as compared with arterioles, but whether this is due to physiological variations in oxygen levels, vessel properties of the veins, or the imaging capabilities of the system is yet unknown.
As this instrument is at its very early stage of development, a number of technical limitations are potential sources of error. First, eye movements are likely to occur during the 10 seconds that are required to obtain the 21 images sequentially. These eye movements may affect the illumination conditions for a given retinal region from one frame to another. Although the general form of the illumination intensity as a function of the position is corrected for by the normalization process, more subtle effects, such as shadowing, and specular and parasitic reflections, may still be present. Furthermore, variations in local oxygen content throughout the cardiac cycle may also occur during the acquisition period. Future optimization of the instrument should result in shorter acquisitions (possibly 1–3 seconds), as frequently observed in optical coherence tomography imaging and limit these effects. Also, it is important to note that the automation of procedures has not yet been implemented within the system. Thus, the determination of OD values is performed manually, which inevitably introduces variability in the process (for example, the line selections performed in the five spectral cubes of a given subject were most probably not from exactly the same location, although effort was made to repeat the selections as closely as practically possible). Furthermore, the image registration algorithms implemented in this preliminary version of the analysis software could not correct for the rotation often observed in the images due to eye movements during data acquisition. Moreover, the perfectible optics design of the retinal camera is responsible for a number of image artifacts that could not be completely corrected in the image-processing step and this may contribute to the data variability observed. However, as image analysis (J-PS) and reflectance extraction and statistical analysis (SRP) was performed by only one user, this did reduce the amount of error that could possibly affect the results.
The next step will be to extract retinal oximetry values from the hyperspectral data. Different approaches have been proposed for this purpose, including curve fitting
5,14 and curve integration.
10,20 To determine the best possible approach, a study is currently under way where hyperpspectral datasets are collected during gas provocation (i.e., that the systemic oximetry is precisely controlled). To identify the best algorithm for retinal oximetry, we will aim to maximize the correlation between the arteriolar retinal oximetry and systemic oximetry values.
Finally, it is pertinent to compare the technology of the HRC based on a TLS used in this study to other hyperspectral instruments described previously. Systems based on push-broom
10,20 and pinhole
14 spectrograph collect spectral-rich information from a line or a point, respectively, of the retina. Reconstitution of a full two-dimensional image is impractical in humans, as the process requires complete immobilization of the eye for the duration of the acquisition, lasting several seconds. In comparison, the system used in our study obtains two-dimensional images for an array of wavelengths. The common landmark features (e.g., optic disk, vessels, macula) can therefore be used to realign the images with theoretically subpixel precision. Furthermore, in contrast to snapshot retinal imaging systems,
11,21,22 where all wavelengths are acquired simultaneously, meaning that the total area of the camera's sensor needs to be separated in the same amount of wavelengths imaged, a bigger field-of-view and/or higher spatial resolution can be achieved with the HRC, as the full sensor surface is used for an image. Last, instruments based on a discrete number of light-emitting diodes
23 and optical filters
12 offer less flexibility and considerably limit the accessible spectral information compared with a tunable laser source.
In summary, this article has described a novel hyperspectral prototype for spectral imaging of the retina that can potentially be used in the future to acquire retinal vessel blood oxygen saturation values. By considering the limitations of ocular imaging encountered by other retinal oximetry studies, namely longer acquisition and exposure times, flash exposure, and limited wavelength intervals, this new instrument may be promising in acquiring more refined and faster measurements of nonflash exposure retinal oximetry measurements in vivo that can potentially be applied to human retinal vascular disease.