Journal of Applied Physiology


A method for noninvasive measurement of Hb O2 saturation (So 2) in retinal blood vessels by digital imaging was developed and tested. Images of vessels were recorded at O2-sensitive and O2-insensitive wavelengths (600 and 569 nm, respectively) by using a modified fundus camera with an image splitter coupled to an 18-bit digital camera. Retinal arterial So 2 was varied experimentally by having subjects breathe mixtures of O2 and N2 while systemic arterial So 2 was monitored with a pulse oximeter. Optical densities (ODs) of vascular segments were determined using a computer algorithm to track the path of reflected light intensity along vessels. During graded hypoxia the OD ratio (ODR = OD600/OD569) bore an inverse linear relationship to systemic So 2. Compensation for the influence of choroidal pigmentation significantly reduced variation in the arterial So 2measurements among subjects. An O2sensitivity of 0.00504 ± 0.00029 (SE) ODR units/%So 2was determined. Retinal venous So 2 at normoxia was 55 ± 3.38% (SE). Breathing 100% O2 increased venous So 2 by 19.2 ± 2.9%. This technique, when combined with blood flow studies in human subjects, will enable the study of retinal O2 utilization under experimental and various disease conditions.

  • hemoglobin oxygen saturation
  • spectrophotometry

the blood supply to the inner retina is well regulated to maintain an adequate O2 supply to this tissue. Despite this circulation’s key role in maintaining viability of the inner retina, little is known about retinal O2 consumption and how it changes in response to disease or other challenges. Noninvasive optical methods for determining the percent blood O2 saturation (So 2) in large retinal vessels in humans and animals have been reported. These methods depend on quantitating differences between oxyhemoglobin (HbO2) and deoxyhemoglobin (Hb) light absorption spectra (16) at different wavelengths. Hickam and co-workers (7, 8) and Laing et al. (12) investigated the use of photographic densitometry of large retinal vessels in vivo to monitor arterial and venous retinal blood So 2. Both groups concluded that accurate measurements could be made using the ratio of vessel optical densities (ODs) at two wavelengths, but their methods depended on external calibration by means of independent arterial So 2measurement. This type of measurement is laborious and inherently limited by the nonlinearity and variable reproducibility of photographic film. These investigators applied the photographic method to study relationships between retinal circulation and arteriovenous So 2differences (7) and assessed its applicability for animal measurements of retinal So 2 (12).

Delori (4) developed a retinal oximetry technique based on photoelectronic measurements of retinal vessel OD at three wavelengths. His method used a mechanical scanner in a modified fundus camera equipped with a photomultiplier detector to monitor sequentially the light reflected inside and outside vessels at three wavelengths. By using the theoretical relationship between OD and light absorption of whole blood described by Anderson and Sekelj (1) and others (11, 15), he was able to avoid the need for external calibration. This method was applied in a study of retinal vein So 2 in optic atrophy (6).

We describe a new technique for the determination of retinal vessel So 2 that uses digitally recorded retinal images obtained simultaneously at two wavelengths: one sensitive to changes in the percentage of HbO2 present in the blood and the other an isosbestic wavelength for HbO2 and Hb. Figure1 shows the effect of differing HbO2 content on reflected light from retinal vessels near the optic disk. Reflected light intensities from arteries and veins at the isosbestic wavelength (569 nm) are comparable, whereas at 600 nm the high percentage of HbO2 increases the reflectance of the artery. Our method utilizes the approximately linear relationship that has been found between Hb So 2 and the OD of hemoglobin in solution (16) and blood (8, 12) as a basis for determining retinal vessel So 2. We use a vessel tracking algorithm to calculate apparent ODs of blood contained in retinal vessels from reflectance measured on the vessel and in the surrounding retinal tissue. Using assumptions about the optical path through the blood and the hemoglobin concentration to determine actual light absorption in the blood, Delori’s method employed similar measurements from vessels to estimate So 2 (4). Our approach differs from those used previously by the use of digital image analysis of reflectance at two wavelengths to determine empirical relationships between ratios of vessel density and external measurements of the systemic So 2. The method is simpler in theory and practice and requires less complex instrumentation than those previously described. It can be adapted to most currently available fundus cameras. We also show that it is possible to compensate for the variability in fundus pigmentation among individuals and races. This technique should be useful for studying diseases that affect retinal O2utilization and for evaluating potential therapeutic interventions.

Fig. 1.

Dual-wavelength image of retinal vessels near optic disk obtained with oximeter. In 600-nm image, arteries (A) appear lighter and veins (V) darker; at 569 nm, both vessel types are dark. Nerve fiber layer can be seen in both images, with greater contrast in 569-nm image.


Study individuals.

Ten healthy nonsmoking men aged 18–40 yr without history of dyshemoglobinemia consented to participate in the study. One subject was unable to complete the oximetry portion of the study. Two subjects were excluded from the oximetry portion of the study, because images obtained from one subject revealed sclerosis of the arterial wall, and in the other subject the artery demonstrated extreme instability of caliber during measurements. Oximetry data from five Caucasians of northern (n = 3) and southern (n = 2) European descent and from two African-Americans were included in the study. Guidelines of the Association for Research in Vision and Ophthalmology for human investigation were followed, and our institution’s Human Investigation Committee approved the study protocol.

Breathing gas-monitoring protocol.

We placed an Ohmeda 3700 (version J) ear oximeter probe on the massaged earlobe. A Mapleson C breathing circuit and face mask were modified by inserting an inspiratory valve just proximal to the mask, allowing all exhaled gas to be vented through the expiratory valve placed just distal to the mask. This arrangement of components kept rebreathing of alveolar gas and dead space to a minimum. A polarographic inspired O2 analyzer with a Clark electrode (Oxychek model 2000, Critikon) was calibrated to room air and inserted just proximal to the inspiratory valve (Fig.2). Subjects were allowed to breathe room air during the first series of fundus image recordings. We then used flows of ≥10 l/min to deliver various O2-N2mixtures (size H gas cylinders). After the initial image recordings during room air breathing, subjects breathed a gas mixture containing 14% O2, which corresponds to an arterial Hb So 2 of ∼94%, or a mixture of 10% O2, which corresponds to an Hb So 2 of ∼84%. In some individuals the 14% O2 mixture was given first, then the 10% O2 mixture. A mixture containing 8% O2 was then given, which corresponds to an Hb So 2 of ∼80%. Finally, subjects breathed 100% O2, which produces an arterial Hb So 2 of 100%. When a steady state was achieved at each step (defined by an unchanging pulse oximeter reading for 2 min), we noted the oximeter saturation reading and obtained retinal images with the fundus camera.

Fig. 2.

Breathing circuit for administering gas mixtures. Velcro straps were used to fix face mask tightly over subject’s mouth and nose. Reservoir bag was inflated with gas mixture before breathing started. FGF, fresh gas flow; IV, inspiratory valve; EV, expiratory valve.

Simultaneous dual-wavelength imaging system.

The retina was illuminated using the internal xenon flash of the fundus camera (model RC/W, Kowa, Tokyo, Japan). Flash energy was supplied by an external strobe power supply (model 404, Norman) at a setting of 200 J, giving a retinal exposure per flash of ∼100 mJ/cm2 (40° field). Flashes were synchronized with the recording system, and wavelengths above and below the recording wavelengths were eliminated with a broad-band filter (580 nm center, 60 nm half-width). The fundus camera was modified to form an intermediate image (II2) just before the exit aperture by using a −6 diopter lens in one of the filter wheel apertures (Fig. 3). A slit placed at this plane contained an ∼1 mm wide × 2 mm high image of the retina. To avoid central light artifacts present in the fundus camera, we positioned the slit off the optical axis. Light forming the intermediate image in the slit was separated into two identical-length paths with image-splitting optics, ultimately forming two laterally displaced images at the solid-state camera detector (Fig.4). The beam was split with a dichroic mirror having a transition at 595 nm (Omega Optical, Brattleboro, NH). Narrow (4.5 nm half-width) band-pass filters with 569- and 600-nm center wavelengths (Omega Optical) were placed in each optical path. We achieved lateral displacement of the images with a series of front surface mirrors that could be moved on rails by means of a microcaliper mechanism. An opaque light baffle blocked stray light between the two images, and a two-element achromatic converging lens (Edmund Scientific, Barrington, NJ) placed before the beam splitter refocused the images at the image sensor. We recorded images with an 18-bit air-cooled digital camera (model OMA, EG & G) equipped with a 1,024 × 1,024-pixel charge-coupled device detector. The camera was controlled using manufacturer-supplied software (HIDRIS). Each image element was formed by 3 × 3 pixels, and the temperature of the charge-coupled device was maintained at −60°C.

Fig. 3.

Optical train of Kowa fundus camera as used with image splitter. A negative lens (DL) contained in 1 filter wheel (W) position produces an intermediate image (II2) at a slit mask (S) that projects past back plate of fundus camera. A second positive lens after slit (CL) refocuses image at camera after passing through optics of image splitter. Xe, xenon flash; F, illumination filter; M, mirrors; A, mirror aperture; OL, ophthalmological lens; VFOL, variable-focus objective lens; IS, image sensor; L, lens.

Fig. 4.

Optical train of image splitter. Beam passing through slit is refocused with lens (L). Beam is split into 2 paths with a dichroic beam splitter (DBS) and redirected using front surface mirrors (M1–M7) to produce laterally displaced images at sensor. Optical path lengths are made equal by adjusting distances of M1 and M2 (left) and M5 and M6 (right) from main optical axis along tracks (T). Images are positioned side by side by on sensor by small adjustments in position of M2 and M6 relative to M1 and M5. A baffle (B) blocks stray light at image sensor from 2 beams. F1 and F2, filters with 569- and 600-nm center wavelengths, respectively; CCD, charge-coupled device.

Measurement wavelengths.

O2-sensitive images were obtained at 600 nm, where light absorption of HbO2 and Hb differ by approximately fourfold (16). A second image was obtained at 569 nm, which is an isosbestic wavelength for Hb and HbO2. Filter half-amplitude bandwidths were 4.5 nm. We employed the isosbestic wavelength, rather than a second O2-sensitive wavelength, to assess stability of vessel measurements.

To determine the effect of filter bandwidth on measurements of blood OD, we calculated apparent light extinction coefficients (ε) of HbO2 and Hb from the products of each absorption spectrum with the filter spectra as described by Delori (4). This product was obtained using polynomial curve fits to each absorption spectrum (Fig. 5, solid lines); two regions of the HbO2 spectrum were fit separately, since it was not possible to obtain an accurate fit over the entire curve. These values were compared with the coefficients of Van Assendelft (16) obtained at high spectral resolution (Table 1). Our filter bandwidth caused measured ODs of fully saturated and desaturated blood to differ by 8% at the isosbestic wavelength. Filter transmission spectra are shown superimposed on absorption spectra in Fig. 5.

Fig. 5.

Points on absorption spectra of oxyhemoglobin (HbO2, •) and deoxyhemoglobin (Hb, •) near measurement wavelengths [data from Van Assendelft (16)]. Solid lines in HbO2spectrum show segments over which curve fits were performed to obtain product of spectrum and filter function. A single curve was fit to deoxyhemoglobin spectrum. Transmission spectra of filters were used to select O2-sensitive (600-nm) and O2-insensitive (569-nm) images (dashed lines). Solid lines, polynomial curve fits. OD, optical density.

View this table:
Table 1.

Comparison of wide- and narrow-band extinction values

Fundus recording.

Eyes were dilated with 1% topical tropicamide and 2.5% phenylephrine. After the pupil diameter stabilized, the room was darkened and the subject’s fundus was illuminated using the tungsten aiming light of the fundus camera. Positioning of eye gaze to enable imaging the retinal area of interest was established using a movable illuminated fixation target against a dark background. Optimal focus was also achieved using the aiming light. The actual data recording at each breathing gas mixture was made using the xenon flash. Three to five measurements were recorded at each gas mixture and stored to disk memory.

Image analysis.

A computer program scanned the digital image to obtain average reflected light intensity from identical vessel segments in both laterally displaced images. Figure 6 shows the path obtained by tracking the minimum reflection inside vessels and the path of the extravascular reflection at a fixed distance from the minimum reflection in both images. To begin a scan, the initialx-y coordinate of the vessel segment was identified in the 569-nm image, where vessel contrast was greater than in the 600-nm image, and x- andy-offsets to the same anatomic point on the 600-nm image were determined. If light reflex artifacts were present in the center of the vessel, the starting point was chosen to lie to the side of the reflex. Between 30 and 120 rows or columns of elements in the 569-nm image were scanned, depending on the orientation of the vessel, to establish the path of the minimum reflected light intensity along the vessel. Edges of the blood column were identified using gradient detection filters (Sobel operators). The edge was defined at pixels for which the sum of filter responses peaked; this procedure did not mistake smaller light gradients of the central vessel reflection for the edge of the blood column. Minimum intensity inside the vessel was then found by scanning between the edges. In arteries, only the minimum-valued element on each line was used; in veins, larger diameter and smaller central reflexes allowed averaging minimum image intensities within a neighborhood of points. The final measurement of light reflection was obtained from the average of minima along the vessel segment. Extravascular light reflection was determined from pixels at a fixed distance outside the vessels. The algorithm could be preprogrammed to average from both sides or only one side of the vessel if strong reflections from nerve fibers were present on one side. We then made similar measurements from the 600-nm image by adding thex- andy-offsets to the vessel path determined from the 569-nm image. Dark pixel values from an area of the detector outside the fundus image were subtracted from measured intensities.

Fig. 6.

Computer-generated scans along retinal vessels in 600- and 569-nm images. Scans inside vessels (thin lines) follow minimum value of reflectance and usually remain to 1 side of central vessel reflections. Scans outside vessels (wide lines) occur at a fixed distance from inner scan. Top and bottom vessels are veins. Middle vessel is an artery. Dark pixel readings are obtained from space between images.

A direct calculation of the apparent vessel OD was obtained at each wavelength from average intensities obtained inside and outside vessels (I in andI out)ODvessel=log10(Iout/Iin) Equation 1 We determined the oximeter response to So 2 from ratios of vessel ODs (ODR) calculated using formulas given below (methods I–III), which compensated differently for choroidal pigmentation.

The degree of retinal pigmentation was assessed from the ratio of extravascular light reflection (EVR) at the two wavelengths used near vesselsEVR=I569/ηI600 Equation 2 where η is the ratio of transmission from the light source to each of the images inherent in the image splitter and camera. We determined η by measuring intensities of the two images reflecting off a neutral reflective surface obtained with barium sulfate paint.

Vessel diameter (D) was measured by determining horizontal and vertical distances (D h andD v) between points where light intensity was midway between values inside and outside the vessel and applying the formula 1/D 2 = 1/ Dh2 + 1/ Dv2 . The image scale was determined from an image of a 0.1-mm ruling.


Stability of recording.

To test the stability of simultaneous reflectance recordings at each wavelength over time, we obtained sets of images while holding systemic arterial So 2 at different values. Figure 7 shows reflected intensities from inside and outside vessels at both wavelengths, which were obtained from images recorded at 3-s intervals in one subject who was told to maintain a fixed eye gaze. At each more hypoxic level, there was greater variation in the reflected light obtained from successive images. This pattern was typical of our recordings and is consistent with our experience that optimal alignment of the eye and eyelid is more difficult to maintain under hypoxic conditions. Because the 569- and 600-nm images were recorded simultaneously, changes in light intensity tended to move in parallel. Vessel ODs averaged from these images for each level of So 2 and corresponding values of ODR obtained using method I (see below) are shown in Table2. Theoretically, OD600 should show an inverse monotonic relationship with So 2, whereas OD569 should remain constant. In practice, these relationships were usually not present in the data for reasons that may be attributable to small changes in focus. However, analyses given below show that when O2-sensitive ODs are normalized by ODs obtained simultaneously at the O2-insensitive wavelength, OD600/OD569does follow an inverse monotonic relationship with the externally recorded systemic So 2. For all subjects in the study, this relationship is fit well by a linear model.

Fig. 7.

Reflected light intensities from images obtained during different levels of systemic O2 saturation (So 2). Symbols (left toright) represent light intensity values for a fixed systemic So 2 from each successive image. Circles and squares, intensities outside and inside vessels, respectively. Open symbols, 600-nm reflectance; filled symbols, 569-nm reflectance.

View this table:
Table 2.

Calculated measures of arterial density at different systemic So 2

Retinal oximetry in arteries.

To characterize the response of our oximeter, we first determined its response in retinal arteries at fixed known values of arterial Hb So 2. We expected that subject-to-subject variations in the amount of pigment contained in extravascular structures would have an effect on our determination of ODs and, therefore, measurements of So 2. Thus we compared our results using three different methods for analyzing the data, each of which treats variations in background pigmentation differently.

Method I: direct reflectance analysis.

In method I, OD600/OD569, determined by direct calculations of vessel OD from each of the subjects, was plotted against arterial systemic So 2 and linear regression analysis was performed between these variables. Slopes of regression lines, which are estimates of O2 sensitivity, were 0.00427–0.01242 [0.00787 ± 0.00098 (SE)] andy-intercepts (ODR at So 2 0%) were 0.514–1.50 (1.02 ± 0.115). The goodness-of-fitR was −0.9581 to −0.9999, with an average of −0.98318 (Table3). These results show that a linear model gives excellent agreement with the change in ODR vs. arterial So 2 between 77 and 100%. Figure8 A shows ODR points obtained from all seven subjects with associated line fits, along with the O2 sensitivity estimated from the average of the slope andy-intercepts. By this method, maximum and minimum slopes varied by a factor of 2.91 and the coefficient of variation in regression slope (SE/mean) was 0.125. We found that standard errors of ODRs were typically <3% of the mean when results from four or more images were averaged.

View this table:
Table 3.

Regression coefficients from retinal arteries

Fig. 8.

OD ratios (ODRs) from retinal arteries vs. systemic So 2 (7 subjects). A: ODR obtained from direct calculation of vessel OD by using method I. B: ODR corrected for retinal pigmentation (ODRcor) by usingmethod II. Dashed lines, line fits through points from each subject; solid line, line obtained by averaging slope and intercept from all subjects: 0.00787 ± 0.00098 (SE) for method I and 0.00504 ± 0.00029 (SE) for method II.C: lines obtained from linear regression between ODR from intravascular reflectance model (ODRirm, method III) and systemic So 2. Maximum and minimum values of slopes, representing O2 sensitivity, are 0.00616 and 0.00437 per unit of So 2 (%), which is approximately the same as range obtained usingmethod II. Vessel diameter indicated for each line fit is expressed in μm (12.33 μm/pixel).

Method II: correction for extravascular pigmentation.

The analysis using method I may underestimate vessel ODs, particularly at 569 nm, because pigmentation in tissue surrounding vessels causes strong light absorption. To assess the effects of pigmentation on light reflectance at our measurement wavelengths, we determined the ratio of EVR at 569 and 600 nm near vessels of the temporal circulation in nine subjects. EVR averaged from five Caucasian subjects was 0.51 ± 0.25. The two lowest values were seen in subjects with blue irises. In one Indian subject, EVR was 0.6, and in three African-American subjects EVR averaged 0.81 ± 0.06. Our values of EVR are in agreement with previous reflectance spectra (5), which showed a greater reduction in reflectance at longer than at shorter wavelengths with increasing pigmentation. When regression slopes obtained by method I are plotted against EVR (Fig. 9, solid symbols), these variables vary inversely (P = 0.057). Reflectance from extravascular retinal tissue was greater at 600 than at 569 nm in all subjects.

Fig. 9.

Relationship between ratio of extravascular reflection at 569 and 600 nm (EVR) and slopes of line fits between ODR and So 2. •, Slopes obtained by direct determination of ODR bymethod I; ○, slopes obtained bymethod II with correction for choroidal pigmentation. AA, African-American; C, Caucasian; BK-BR, black hair and brown eyes (refers to 3 neighboring points); BR-BL, brown hair and blue eyes; R-BL, red hair and blue eyes.

To reduce influences of pigmentation, we applied a second formula for the vessel OD at 569 nm that used light reflected outside the vessel at 600 nm to estimate incident light at 569 nmODcor,569=log10(ηIout,600/Iin,569) Equation 3 where η is defined in Eq. 2.

An ODR corrected for pigment effects (ODRcor) was calculated asODRcor=OD600/ODcor,569 Equation 4When values of ODRcor from all subjects were plotted against arterial So 2, the variance of linear regression coefficients between subjects was significantly reduced with respect to those obtained using the uncorrected ODR. Slopes were 0.00433–0.00633 (0.00504 ± 0.00029, coefficient of variation = 0.058), andy-intercepts were 0.492–0.775 (0.617 ± 0.033). The correction did not significantly change the standard errors of points for individuals. Figure8 B shows ODRcor points from the same seven subjects with line fits, along with the O2 sensitivity averaged from individual slopes and intercepts. Results of regression analysis usingmethod II are given in Table 3 along with uncorrected results. A plot of slopes obtained bymethod II against EVR (Fig. 9, open symbols) showed significantly less correlation with EVR (P = 0.19). These results indicate that light absorption from retinal pigments, which is especially strong near 569 nm, influences determinations of vessel So 2 from reflectance measurements and show that a correction for pigment variation, based on estimation of incident illumination at both wavelengths using values measured at 600 nm, provides a twofold reduction in the variability of O2sensitivity among subjects obtained by our method.

To determine whether the range in O2 sensitivity remaining after pigment correction could have resulted from different vessel sizes (8), a second linear regression of line-fit parameters against vessel diameter was performed. Significance (P < 0.05) was obtained for values of ODRcor obtained while subjects breathed 100% O2 (Fig.10,top). There was a positive trend between the slope of the regression and diameter that was not statistically significant (P = 0.21). These results suggest that the range in ODRcor values that was associated with a given value of arterial So 2 was partially the result of different vessel sizes and that points in this range can be correlated to a measurement obtained under repeatable conditions (100% O2 breathing).

Fig. 10.

Relationship between measurement parameters and vessel diameter.Top: ODRcor at 100% O2 (method II) vs. diameter in pixels. Linear regression slope = 0.0129/pixel (R = 0.7739,P = 0.045).Bottom: slope of regression between ODRirm and systemic So 2(method III) vs. diameter. Linear regression slope = 18.4/pixel (R = 0.8224, P = 0.023).

Method III: intravascular reflectance model.

An estimate of So 2 that is independent of effects of extravascular pigmentation can be made by comparing reflectance from only within blood vessels at the two wavelengths and replacing measured reflectance outside vessels with an estimate of reflectance from unpigmented fundus. If arterial density at 100% So 2is set primarily by light absorption by HbO2, an appropriate estimate of unpigmented reflectance (UR) would set vessel ODs at each wavelength to values predicted from the HbO2absorption spectrum. We propose a model that yields this value of UR from Hb extinction coefficients (broad-band coefficients in Table 1) and measured reflectance from inside vessels at each wavelength when blood is 100% saturated with O2(intravascular reflectance model)εHbO2,600/εHbO2,569=ODRirm,100% =log(UR/Iin,600,100%)/ Equation 5 log(UR/Iin,569,100%) =0.0797(εvalues from Table1) We used Eq. 5 with measurements ofI in obtained from images recorded during 100% O2breathing to calculate UR by an iterative method. We then applied this value of UR to measurements during normoxic and hypoxic recordings of each subject to calculate ODR derived by the intravascular reflectance model (ODRirm) at each breathing set point. Extravascular intensity scanned near the vessel at 600 nm was used to compensate for differences in illumination at each point. Figure 8 C shows regression lines that were fit to these sets of points for each subject, along with diameters of each artery. Initial assumptions cause the lines to meet near 100% So 2, and thus this plot shows more clearly the range in slopes from our subjects. It was not possible to predict slopes from subject coloration or from the EVR measured in each subject. In this model, regression intercepts and slopes are codependent; correlation between the slope and vessel diameter was significant to the 0.05 level (Fig. 10,bottom). These results again suggest that there is a small dependence of the O2 sensitivity on size of vessels in the range 80–200 μm.

Venous So2.

Figure 11 shows ODRs of an artery-and-vein pair recorded from the temporal retinal circulation in one subject. Although a priori there is no reason to expect that venous So 2should vary linearly over the range of arterial saturations studied, we found that venous ODRs obtained during room air breathing and hypoxia fell on straight lines, whereas the ODR at 100% O2 breathing was below the line, a finding similar to that reported by Hickam and Frayser (7). Presumably, the high concentration of dissolved O2 at 100% significantly reduces dissociation of HbO2 during capillary passage, lowering the ODR. This effect was seen even though vasoconstriction during hyperoxia reduces blood flow, allowing more extraction of O2, which, by itself, would have increased the ODR.

Fig. 11.

Values of ODRcor obtained for an artery-vein pair plotted against systemic So 2. Vein slope = −0.00315/%So 2(R = 0.994, excluding 100% O2 point). Artery slope = −0.00808/%So 2(R = 0.992). Error bars, SE.

We used the calibration obtained by method II to estimate So 2 in segments of retinal veins (5 subjects) that were near arteries recorded during calibration experiments. To evaluate venous So 2 at an experimental O2 set point, we first determined an equivalent venous ODRcor that would have been obtained if the vein and paired artery had equal diameters. This new value is equal to the measured venous ODRcor minus the product of the slope of arterial ODR at 100% O2vs. vessel diameter (Fig. 10,bottom) and the difference in diameter of the artery and vein. So 2 was then determined by subtracting the adjusted venous ODRcor from the arterial ODRcor obtained at 100% O2 and dividing this quantity by the mean value of O2 sensitivity found from the pigment-corrected calibration in arteries (0.00504 ODR unit/%So 2). Venous So 2was then found by subtracting this value from 100%. This calculation is summarized as followsSvO2=100% Equation 6 [ODRcor,arteryODRcor,vein(ΔD×C)]/OS where SvO2 is venous So 2, ΔD is the difference in diameters of the vein and paired artery, C is the slope of the line fit between arterial ODRcor at 100% O2 and diameter (0.0129/pixel), and OS is O2 sensitivity (0.00504/%So 2). Results of this analysis are summarized in Table4 for room air breathing by each subject. Venous So 2was 55 ± 3.39% (SE) with a range of 26% among subjects. The increase in venous So 2 caused by breathing 100% O2 (hyperoxia relative to normoxia) was determined by subtracting the normoxic value of ODRcor from the hyperoxic value and dividing by OS. In this case, size-dependent offsets in the vessel ODR were cancelled by subtraction. We found that hyperoxia (100% O2 breathing) increased venous So 2 by 19.2 ± 2.9%, with the smallest increase in the subject with the highest saturation at rest (Table 4).

View this table:
Table 4.

Venous So 2


We have described a new imaging method for measuring retinal vessel Hb So 2. The use of an image splitter and high-resolution camera to image vessels simultaneously at two wavelengths is an extension of the earlier work of Hickam et al. (7, 8) using photographic film and offers significant advantages. Almost perfect linearity and reproducibility of the solid-state camera enable detection of small differences in reflected light associated with changes in vessel So 2. Equally important, we found that a ratiometric approach effectively cancels effects of subject motion, which otherwise interferes with the recording process. We used the image recorded at 569 nm as a reference image with which to cancel changes unrelated to blood oxygenation in the 600-nm image. With these recording techniques it was possible to obtain estimates of blood OD by relatively simple vessel-tracking methods.

We found in each subject that the ODRs, when plotted against 77–100% systemic So 2, fell along straight lines, with R close to −1. These results agree with previous ratio measurements obtained from retinal vessels by reflectance (8, 12) and are predicted over our range of So 2 by assuming the quasi-linear D-E characteristic described by Delori (4). Using reflectance pulse oximetry, de Kock et al. (3) showed a nonlinear relationship at <85% saturation in vitro; however, the method of de Kock et al. relied on pulsatile changes in reflectance from vessels and surrounding retinal tissue to detect arterial blood, rather than directly imaging vessels to determine apparent ODs.

Differences in choroidal pigmentation among subjects initially influenced the sensitivity of our oximeter to changes in retinal vessel So 2, producing variability in slopes of line fits against known systemic So 2 values. We were able to reduce the effect of light absorption by extravascular pigment by substituting reflectance at 600 nm, where pigment light absorption is substantially weaker than at 569 nm (method II). This correction significantly reduced variability in regression coefficients. Elimination of effects of extravascular pigment by using only measures of intravascular reflectance (method III) did not result in significant further reduction in the variation of O2 sensitivity among subjects. Neither method II norIII addressed effects of pigment behind the vessel that may still exert an influence on the apparent vessel OD. At longer wavelengths, including 600 nm, retinal reflectance has been shown to be more variable than at shorter wavelengths among subjects of different coloration (5). However, we used our longer wavelength to correct for the influence of choroidal pigmentation, because the degree of light absorption at 569 nm by these pigments is significantly greater than at 600 nm. The stronger light absorption at 569 nm produces a larger numerical effect on the calculation of vessel OD than do variations at 600 nm.

After correction for pigmentation, remaining scatter in ODR parameters was correlated with vessel diameter. Hickam et al. (8) and Delori (4) found that their measures of So 2 were dependent on vessel size: Hickam et al. concluded that similarly sized vessels near the optic disk could be accurately measured with a single calibration, and Delori incorporated vessel diameters into his theoretical analysis. We adjusted measured ODR values for vessel size. Further improvement in accuracy of our technique by usingmethod II possibly could be obtained if vessel size dependence of O2sensitivity (i.e., effect on slope) were also included. However, the degree of correlation between the regression slope for ODRcor (method II) and diameter in our data was not statistically significant; hence, we have used the mean value of slopes to define O2 sensitivity. We found withmethod II that the vessel ODRcor obtained at 100% O2 is dependent (P < 0.05) on vessel diameter, and it is easy to measure. If this parameter is obtained, retinal arterial So 2 can be readily determined by our method. A statistically significant correlation did exist between slopes obtained by our intravascular reflectance model (method III) and systemic So 2; however, this second method required more computation. Becausemethods II andIII reduced effects of pigmentation on O2 sensitivity by comparable amounts and because the simpler method II yielded reasonable estimates of venous So 2 in veins of different diameter, we believe that method II is most suitable for oximetry. The results with both methods indicate that vessel diameter has a slight influence on measurement of vessel blood So 2 and that compensation for the effect of pigmentation is necessary before effects of vessel size can be addressed.

We have based our determinations of vessel So 2 on the O2 sensitivity found from means of regression slopes in arteries, with correction for vessel size using the diameter dependence of arterial ODR at 100% O2. Determination of the actual O2 responses in arteries during graded hypoxia is not practical in a clinical setting or each time a new measurement is needed. There has also been no practical way to independently validate retinal venous saturation measurements in human subjects, making it difficult to interpret O2 extraction data. Confidence in the accuracy of venous So 2measurements relies on agreement reached using different instrumental and analytic methods. The mean venous saturation determined by our method is in general agreement with values obtained by the three-wavelength method of Delori et al. (4, 6) and in close agreement with results from the earlier work of Hickam et al. (7, 8) using dual-wavelength ratio analysis from photographic images. Hickam et al. and we assumed that the arterial calibration between ODR and externally measured systemic saturation could be applied to venous ratios by extrapolation toward values of lower saturation. This assumption requires that the relationship be linear over the range of saturation found in arteries and in veins. Shortly after Hickam et al. reported their measurements of venous So 2 in human subjects, Laing et al. (12) published their “falling saturation” experiment in the rabbit, confirming that retinal arterial ODRs vary linearly with blood So 2 between 20 and 97%, which covers the normal saturation range of both vessel types. The practical problem with use of arterial calibration data for venous measurement is that any error in the best-fitting regression line leads to inaccurate estimates outside the range of the calibration. Our determination of venous saturation is subject to this kind of error. We note, however, that we were able to obtain goodness-of-fit values close to −1 (−0.94 to −0.99) in regression line fits for all our subjects, which increases confidence that a dual-wavelength method yields the expected linear relationship between ODR and So 2 and that the slope of the best-fit line is a good predictor of other points outside the range of calibration.

Verification of our oximeter response depends on the accuracy of external So 2measurements. Clinical studies have demonstrated that the accuracy of pulse oximeters under steady-state conditions is within ±2 saturation points within 100 to 70% saturation (2, 9, 13). We chose to use an ear probe, rather than a finger probe, because of its ability to respond accurately and quickly to a changing systemic arterial saturation. Kagle et al. (9) demonstrated that the Ohmeda 3700 version J ear probe is highly accurate during hypoxic conditions in normal volunteers, giving a goodness-of-fit Rof 0.98 for regression of ear probe readings vs. arterial saturation. The slope and intercept for the African-American subjects did not differ significantly from that for the Caucasian subjects, a finding also described by Cecil et al. (2) using the Ohmeda 3700 finger probe. One of the most significant limitations of the pulse oximeter is signal failure due to poor perfusion or hypothermia. The Ohmeda 3700 features a graphic display of waveform and signal strength to aid in minimizing inaccurate readings. We monitored the quality of ear probe signals during our 2-min stabilization period with steady-state oxygenation conditions before recording oximeter readings and obtaining retinal images and used only readings with a strong signal.

Sources of possible error by our technique include effects of changes in red cell aggregation and orientation on light transmission, which was observed at different blood shear rates by Klose et al. (10). This effect of flow must be considered in evaluating oximetry measurements by reflectance. We note that for light transmission of the blood at our measuring wavelengths to change by amounts necessary to affect our results by using a ratio analysis, shear stress would need to change by over two orders of magnitude. This degree is greater than that associated with autoregulatory changes in the larger vessels. Errors could also arise if either light absorption in extravascular pigment or the efficiency of light scattering from blood and surrounding tissue were to change between measurements. We assume that these effects remain constant at fixed sites and thus will not influence measurements of So 2 when identical vessel segments are imaged before and after changes. The small amount of contrast of arteries against surrounding tissue at 600 nm, which is consistent with the low light absorption by HbO2, suggests that there is not a strong difference in amounts of light scattered by blood and surrounding tissue backward to the fundus camera, which could also contribute to the vessel OD. Although light absorption appears to govern the changes in vessel OD in our images, determinations of the actual blood OD are complicated by light scattering within the blood and by multiple angles of incident light (4), and thus our calculations by Eq. 1 give only an apparent OD for the vessel. Our analytic approach avoids assumptions about the optical path through the vessel, relying instead on empirical relationships between image parameters and vessel So 2obtained from subjects of different coloration. This approach does give reasonable estimates of Hb So 2 in the larger retinal arteries and veins and should be applicable to 50- to 200-μm vessels.

Perhaps more interesting for understanding effects of metabolism or blood flow changes on retinal function is the ability to determine the changes in So 2 before and after interventions (14). The change is more straightforward to obtain, since many variables potentially affecting absolute measures of So 2 will cancel. However, error could occur if vessel diameter changed significantly between measurements. The vessel image should be checked in each recording to ensure that the vessel remains well focused and that diameter changes do not occur. If they do, then the change in diameter should be included in the analysis. The relatively strong vasoconstriction response during breathing pure O2 causes the larger retinal vessels to constrict by <15% (7), an amount that does not seriously effect the O2 sensitivity of our method. We conclude that Hb So 2 in retinal vessels and arteriovenous differences in So 2 can be determined from simultaneous dual-wavelength measurements. The technique may find application in studying effects of experimental interventions or treatments on retinal O2 utilization.


The authors gratefully acknowledge private donation support from James E. Garrette and an unrestricted grant from Research to Prevent Blindness.


  • Portions of this study were presented at the Annual Meeting of the Association for Research in Vision and Ophthalmology, 1997 and 1998.

  • The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked “advertisement” in accordance with 18 U.S.C. §1734 solely to indicate this fact.


View Abstract