2.1.

Description of the System

The developed MSI fundus camera was designed using the Zemax optical simulation software (Radiant Zemax LLC). Customized illumination and detection strategies were conceptualized to achieve, through the whole working wavelength range, the best light collection efficiency, aberration balance and cost (only commercial optical components were used), the most uniform field of illumination over the retina, and the minimal presence of ghost images. The resulting system (Fig. 1) has two detection arms: one for the spectral range from 400 to 950 nm (VIS–NIR detection arm) and another for 960 to 1300 nm (NIR detection arm). Both arms share the objective lens, placed in front of the patient’s eye. Regarding the illumination arm, custom-made rings of LEDs were used rather than the most traditional approach of common fundus cameras, where a mask is inserted in the optical path to create a continuous illumination ring on the pupil. Specifically, 32 surface-mount device LEDs were welded in a ring-shaped configuration on three printed circuits board, each responsible for the illumination in a narrowband (Figs. 1 and 2): the VIS ring (2a), with 18 LEDs (9 twin pairs arranged symmetrically) emitting from 400 to 700 nm; the NIR ring 1 (2b), with 10 LEDs (5 twin pairs) emitting from 700 to 1100 nm; and the NIR ring 2 (2c), with 4 LEDs with a peak wavelength of 1213 nm. The former 15 LED clusters have peak wavelengths of 416, 450, 471, 494, 524, 595, 598, 624, 660, 732, 865, 955, 1025, 1096, and 1213 nm. The light from the rings, attached to a custom-made aluminum structure (1), is then combined in a common optical path by means of two dichroic mirrors with cutoff wavelengths at 1150 nm (3) and 700 nm (4).

Fig. 1

(a) View of the multispectral fundus camera. (b) Optical scheme. Illumination arm: (1) LED rings structure; (2a), (2b), (2c) LED rings; (3), (4) 1150- and 700-nm dichroic mirrors; (5a), (5b) illumination telescopes’ lenses; (6) retinal conjugated diaphragm; (7) antibackreflections stops; (8) fixation target; (9) mirror; (10) beam splitter plate. Detection arms: (11), (12) objective’s lenses; (13) holed mirror; (14) lens; (15) 950-nm dichroic mirror. VIS–NIR detection path: (16) lens; (17) field diaphragm; (18a), (18b) VIS–NIR telescope’s lenses; (19) refraction correction translator platform; (20) aperture diaphragm; (21) CMOS camera. NIR detection path: (22) field diaphragm; (23) mirror; (24) platform-linear translator; (25a), (25b) NIR telescope’s lenses; (26) aperture diaphragm; (27) InGaAs camera.

Fig. 2

Picture of the LED rings and normalized intensity.

After some lenses, mirrors and stops that prevent the back reflected light from the optical surfaces entering inside the detection optical path (5 to 7), a 45-deg custom-made holed mirror (13) conjugated with the patient’s pupil plane couples the illumination (reflected) and detection (transmitted through the mirror’s central hole) paths. This configuration assesses the geometrical pupil separation required to avoid ghost images caused by reflections (from the cornea and lens) of the illumination in the acquired fundus images. Finally, the LED rings are imaged by the objective lens (11 and 12). A properly designed set of lenses (14, 16, and 18 in the VIS–NIR; 14 and 25 in the NIR) is placed in the detection paths after the holed mirror in order to image the retina in the two cameras (21 and 27), and to compensate for the patients’ spherical refraction (range: ± 15 D) (19). The split of the VIS–NIR and the NIR detection paths is achieved with a dichroic mirror (15) with a cutoff wavelength of 950 nm. Field (17 and 22) and aperture (20 and 26) diaphragms are also used. The camera sensors are a CMOS detector (Orca Flash 4.0, Hamamatsu, Japan), with $2048\times 2048\text{pixels}$ and 16-bit depth, and an InGaAs camera (C12741-03, Hamamatsu, Japan), with $640\times 512\text{pixels}$ and 14-bit depth, for the VIS–NIR and NIR detection arms, respectively. Finally, a fixation target (8) is placed at a retinal conjugate plane in the illumination arm to achieve adequate patient fixation while performing measurements. All the optical components of the system were carefully selected so that their working spectral range matched the working spectral range of the arm they are part of, in order to minimize the undesired reflections and maximize the transmittance. In particular, some components, such as the objective lens elements, must work for both the VIS and NIR ranges. In those cases, components with a broadband antireflection coating were chosen, ensuring a reflectance of $<1.25\%$ between 400 and 1000 nm according to the manufacturer’s specifications (Edmund Optics Inc). For longer wavelengths, this reflectance tends to increase (4%), although it did not affect significantly the overall performance of the system.

LEDs and cameras are synchronized by means of trigger signals in the electronics, allowing the sequential recording of the whole set of 15 spectral images in 612 ms [264 ms for the CMOS (VIS–NIR) and 348 ms for the InGaAs (NIR)], thus avoiding the imaging problems caused by eye movements. A user-friendly MATLAB interface was developed to control the system’s operation and to set the parameters.

The resulting MSI prototype, built in accordance with ISO standards for ophthalmic instruments and fundus cameras (ISO 15004-2:2007²² and ISO 10940:2009²³), has an angular field of view of 30 deg, magnifications of 1.20 (VIS–NIR) and 0.88 (NIR), and a resolution better than the limit imposed by the imaging sensors (pixel size 6.5 and $20\mu \mathrm{m}$ for the CMOS and the InGaAs, respectively). The stability of the system was tested through a mid-term analysis by performing 10 measurements in 5 consecutive days, early in the morning and late in the afternoon, and short-term repeatability (10 complete streams of images taken consecutively) using an artificial eye (OEMI-7, Ocular Instrument Inc.). Averaged digital levels on the same portion of the image were then calculated for all wavelengths and measurements; percentages of variation (standard deviation/average) for both the mid- and short-term analyses were found to be under 2.98% and 1.98%, respectively. Since the LEDs do not work with a closed-loop intensity control, periodic calibrations of the system must be carried out to guarantee an accurate retrieval of the spectral features of the retina at all wavelengths along all LEDs lifecycle.⁹ Finally, after the light hazard protection evaluation, the MSI camera was classified as group 1 (nonhazardous ophthalmic instrument).

2.2.

Selection of Wavelengths

In addition to commercial availability, optical characteristics such as radiometric power, angle of emission and dimensions, the wavelengths of the LEDs (Fig. 2 and Table 1) were carefully selected to obtain the most accurate information from the spectral characteristics of structures and biological substances present in the ocular fundus.^10,24 The LEDs from 400 to 500 nm were chosen to investigate the superficial layers of the retina, in particular the nerve fiber (NF) layer and internal limiting membrane, and also the fluorescent retinal components (carotenoids of photoreceptor pigments acting as a yellowish filter in the macula, and lipofuscin that accumulates as a product of phagocytosis of photoreceptors in the outer segments and concentrates in the RPE). Similar to the conventional red-free images obtained with a green filter in commercial fundus cameras, wavelengths within the range of 500 and 700 nm were chosen to study the retinal vascular layers and to discriminate between arteries and veins (oxygenated and deoxygenated blood). It can be noticed that two LEDs with similar peak wavelengths (595 and 598 nm) were included in the system (Table 1), being the spectral width of the latter (76 nm) much larger than for the rest of LEDs in the VIS range. This LED was chosen to sample part of the 530 to 580 nm range, where narrow spectral width LEDs with enough radiometric power and reduced size to be integrated in the system were not commercially available, thus avoiding this gap in the spectral sampling. Melanin absorbs in the spectrum used by traditional retinographs, but in the range from 450 to 600 nm, it is obscured by other dominant ocular components such as hemoglobin, lutein, and rhodopsin.¹² Beyond 600 nm, melanin is the dominant retinal pigment and thus, images at these wavelengths enhance the visualization of retinal structures that contain this pigment, such as the underlying RPE. In addition to deeper penetration in the retina due to the lower absorption of melanin, LEDs with peaks at 732 and 865 nm were expected to provide information about the oxygenation of the deepest retinal and superficial choroidal vasculature (CV). The longest wavelength LEDs (peaks from 955 to 1213 nm) were selected to study the choroid, since the retina appears progressively more transparent, even projecting light into the sclera, which can be used for imaging congenital hypertrophy of the RPE, choroidal ruptures, and melanomas.¹² Lipids (drusen) also show absorption peaks in the NIR that exceed water absorption, and therefore images taken at longer wavelengths can be relevant in the study of AMD.²⁵

Table 1

Peak wavelengths, full width at half maximum (FWHM) and power of the LEDs, and exposure times of the cameras used in the multispectral acquisition for each of them. These times are optimal to compensate for the different powers of the LEDs, the transmittance/reflectance of the optical components of the system and the eye, and the spectral sensitivity of the cameras.

2.3.

Fundus Reflectance Retrieval

To compute reflectance values at each pixel ($i,j$) for a given wavelength (${\lambda }_n$) from the spectral images recorded, we acquired images from a calibrated reference white (BN-R98-SQC, Gigahertz-Optik GmbH, Germany) to compensate for the differences in power of the LEDs and sensitivity of the cameras along the spectral range. Dark current images taken without ambient light and any eye in front of the system were also recorded immediately after the acquisition of each multispectral sequence of the ocular fundus, as a measurement of the background noise. Ocular fundus, reference white, and dark current images were captured using the same exposure times for each spectral band (Table 1). The reflectance was then calculated as commonly described in the literature:^9,10,26