Automatic Retinal Oximetry

Abstract

purpose. To measure hemoglobin oxygen saturation (SO₂) in retinal vessels and to test the reproducibility and sensitivity of an automatic spectrophotometric oximeter.

methods. Specialized software automatically identifies the retinal blood vessels on fundus images, which are obtained with four different wavelengths of light. The software calculates optical density ratios (ODRs) for each vessel. The reproducibility was evaluated by analyzing five repeated measurements of the same vessels. A linear relationship between SO₂ and ODR was assumed and a linear model derived. After calibration, reproducibility and sensitivity were calculated in terms of SO₂. Systemic hyperoxia (n = 16) was induced in healthy volunteers by changing the O₂ concentration in inhaled air from 21% to 100%.

results. The automatic software enhanced reproducibility, and the mean SD for repeated measurements was 3.7% for arterioles and 5.3% venules, in terms of percentage of SO₂ (five repeats, 10 individuals). The model derived for calibration was SO₂ = 125 − 142 · ODR. The arterial SO₂ measured 96% ± 9% (mean ± SD) during normoxia and 101% ± 8% during hyperoxia (n = 16). The difference between normoxia and hyperoxia was significant (P = 0.0027, paired t-test). Corresponding numbers for venules were 55% ± 14% and 78% ± 15% (P < 0.0001). SO₂ is displayed as a pseudocolor map drawn on fundus images.

conclusions. The retinal oximeter is reliable, easy to use, and sensitive to changes in SO₂ when concentration of O₂ in inhaled air is changed.

A noninvasive device to measure retinal oxygen levels is needed for the study of the role of oxygen levels in human retinopathy and glaucoma. Spectrophotometric measurements of hemoglobin oxygen saturation in human retinal blood vessels was initiated by Hickam et al. ¹ in the 1960s but has still not been mastered as a reliable and useful clinical or research tool. During the past 7 years, we have built on the work of Beach et al., ² Delori, ³ and others ⁴ to develop a noninvasive automatic retinal oximeter, which we hope will allow studies of human retinal and optic nerve disease in research and clinical settings. We report the sensitivity and reproducibility of this device in human studies. 

Michaelson (1948) ⁵ suggested that hypoxia plays a role in the pathophysiology of diabetic and other retinopathies. Animal research and limited studies on humans suggest that oxygen may be an important factor in the pathogenesis and treatment of some retinal diseases. These include venular ^{6 7 8 9 10 11} (Scibor M et al. IOVS 2002;43:ARVO E-Abstract 3305) and arteriolar ^{12 13} occlusions, retinal detachment, ^{14 15 16 17 18 19} and diabetic retinopathy ^{20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35} (Trick GL et al. IOVS 2005;46:ARVO E-abstract 1413). Animal studies have also indicated that both vitrectomy ^{7 36 37 38} and laser photocoagulation ^{8 9 28 36 38 39 40 41 42 43 44 45} raise the oxygen tension in the retina. 

Although glaucoma therapy is currently aimed at lowering intraocular pressure, the beneficial effects may be due to better perfusion and oxygenation of the optic nerve rather than the direct effects of lower intraocular pressure. Experiments in monkeys, ^{46 47} cats, ^{48 49} rabbits, ⁵⁰ and pigs ^{51 52} have demonstrated that optic nerve oxygenation decreases when the perfusion pressure declines and autoregulation of blood flow is overwhelmed (e.g., because of raised intraocular pressure). Poor regulation of blood flow has been postulated as the mechanism of glaucomatous optic atrophy. ^{53 54}  

These theories are predominantly based on animal research, and few studies have been conducted to investigate oxygenation in humans, because reliable and noninvasive methods of measuring human retinal oxygenation have been lacking. In the past decades, various approaches to retinal oximetry have been tried (for a review of the methods, see Harris et al. ⁴ ). Most of these methods are based on the fact that oxygenated and deoxygenated hemoglobin have different light absorption spectra (color). By analyzing the light absorbance of blood at two or more wavelengths, the oxygenation of hemoglobin can be estimated. Investigators in this field have run into numerous problems, however, that have prevented the development of a reliable and practical method. The problems include nonlinear sensors (photographic film), eye movement during lengthy measurements, small areas measured at one time, and optical complexities within the eye. Nevertheless, considerable progress has been made. ^{2 3 55 56}  

In this study we compared the reproducibility of our manual and automatic methods for retinal oximetry. We present calibration for our automatic method and use this to display the reproducibility and sensitivity of our retinal oximeter to hemoglobin oxygen saturation. 

Methods

All procedures were approved by the institutional review boards of the respective institutions: Indiana University-Purdue University, Indianapolis, and the University of Iceland, National University Hospital, Reykjavík. Informed consent was provided by all subjects before participation in the studies. All procedures conformed to the tenets of the Declaration of Helsinki. 

The Retinal Oximeter

Our method for retinal oximetry is based on the work of Beach et al. ² The setup and outcome are shown in Figure 1 . 

Two retinal oximeters were used for the study, one based in Indianapolis and the other in Reykjavik. Both are composed of a fundus camera coupled with a beam splitter to a digital camera. (In Indianapolis, the fundus camera was a Topcon TRC50-VT; Topcon Co., Tokyo, Japan; the beam splitter was a MultiSpec Patho-Imager; Optical Insights, Tucson, AZ; and the digital camera was a CoolSNAP ES; Roper Scientific, Tucson, AZ. In Reykjavik, the same beam splitter was used, but the fundus camera was a Canon CR6-45NM; Canon Inc., Tokyo, Japan; the digital camera was an SBIG ST-7E; Santa Barbara Instrument Group, Santa Barbara, CA, used at 2 × 2 binning). The beam splitter separates the original image into four optical channels. In each channel, there is a different narrow band-pass filter, through which only light of specific wavelengths can pass. The center wavelengths of the filters are 542, 558, 586, and 605 nm and the half-bandwidth is 5 nm, except for the 542-nm filter, which has a 9-nm half-bandwidth. In the present study, only the 586- and 605-nm wavelengths were used. 

With a specialized computer program, optical density (OD) can be calculated for every vessel at each of the four wavelengths. OD is a measure of the blood’s light absorbance and is calculated as  

\[\mathrm{OD}{=}\mathrm{log}(\ \frac{I_{0}}{I}),\]

where I and I _o are the brightness levels inside and just outside a vessel, respectively. It can be shown that the ratio of ODs at certain wavelengths (OD ratio, ODR) has an inverse and approximately linear relationship to hemoglobin oxygenation ^{2 4} :  

\[\mathrm{SO}_{2}{=}a{+}k{\cdot}(\ \frac{\mathrm{OD}_{X}}{\mathrm{OD}_{Y}}){=}a{+}k{\cdot}\mathrm{ODR}.\]

In equation 2 , SO₂ is the percentage of hemoglobin oxygen saturation; a and k are constants; OD _X and OD _Y are ODs (no unit) at wavelengths X and Y, respectively; and ODR is the optical density ratio. Thus, in theory, hemoglobin oxygenation can be calculated using brightness inside and outside vessels at two wavelengths of light. 

The Software

Two versions of software were used. Both versions register all four output images in the same coordinate system, so that each pair of points, used for calculation of ODs, is at the same fundus location for all four images. The registration algorithm utilizes the edges of the four images. ⁵⁷ The older software version requires the user to select measurement points manually, inside and outside vessels to calculate the ODR. The newer version automatically locates vessels on the fundus image and selects measurement points inside and outside vessels. Morphologic filters are used to enhance vessellike structures in the images. After morphologic enhancement, a skeleton of a thresholded fundus image is used to locate the approximate centerlines of the arterioles and venules. To avoid error caused by specular reflection, the new software uses the darkest point on the vessel’s cross section. To reduce variation in calculated ODRs due to random system noise and variable optical properties of the fundus, the software averages measurements from several pixels adjacent to the vessel. When the user has specified the vessel segment of interest, the mean ODR for that segment is given. The automatic software also displays a color-coded map of the vessels, where the colors indicate hemoglobin oxygen saturation (after calibration, described later). 

Reproducibility

In Reykjavik, five images (each composed of four spectral images) were taken of one eye of each of 11 individuals. One individual was excluded from analysis because of poor-quality images. The images used were of diabetics, with (n = 2) and without (n = 3) retinopathy, and healthy volunteers (n = 5). The reproducibility was very similar in diabetics and healthy individuals. One first-degree arteriole and one first-degree venule were chosen for analysis in each eye. Each vessel was analyzed with both software versions and the SD was calculated between the five images. When the manual software was used, four pairs of measurement points were averaged for each vessel. 

Sensitivity

The sensitivity of the system was tested by comparing results obtained during inhalation of different concentrations of O₂. In Reykjavik, five healthy volunteers inhaled either 21% (room air) or 100% O₂. In Indianapolis, 15 healthy volunteers inhaled either room air or 100% O₂. Finger pulse oximeters were used to monitor arterial oxygenation (Medical/GE Healthcare, Milwaukee WI in Reykjavík, and POET II model 602-3; Criticare Systems, Waukesha, WI in Indianapolis). During normoxia, the finger oximetry values were 97% to 99% but 98% to 100% during hyperoxia. Measurements were made on one first-degree arteriole and one first-degree venule in each eye. Individuals were excluded from further analysis if their images did not contain the same vessels during inhalation of different concentrations of oxygen (one individual in Reykjavik and three in Indianapolis). A paired t-test was used for statistical analysis. 

Results

Reproducibility

Table 1shows the mean and SD of the five repeated measurements of the same vessel, using either the manual or the automatic software. All other results (below) were generated with the automatic software. 

Calibration

Earlier measurements on healthy volunteers (for review, see Harris et al. ⁴ ) have yielded retinal arteriolar oxygenation of 92%, ⁵⁸ 97%, ⁵⁹ and 98% ³ and venular oxygenation of 45%, ³ 55%, ² 58%, ⁵⁸ and 58%. ¹ The mean of these results is 96% for arterioles and 54% for venules. From our data, we calculated a mean ODR of 0.209 for arterioles and 0.502 for venules (18 healthy individuals, one first-degree vessel pair in each). Assuming a linear relationship between ODR and SO₂ and using the above means from the literature for SO₂ and our mean ODR gives the following:  

\[96\%{=}a{+}k{\cdot}0.209\]

and  

\[54\%{=}a{+}k{\cdot}0.502.\]

Solving equations 3 and 4 4together gives a = 125 and k = 142, and so the equation for SO₂ is  

\[\mathrm{SO}_{2}{=}125{-}142{\cdot}\mathrm{ODR}.\]

 

Reproducibility of Calibrated Results

Table 2shows the reproducibility of the automatic measurements in Table 1after calibration with equation 5 . 

Oxygen Sensitivity of the Automatic Approach

Four healthy volunteers in Reykjavik and 12 in Indianapolis inhaled either room air or 100% oxygen. Table 3shows the results of analysis with the automatic software, using equation 5 . 

There is a significant difference between normoxia and hyperoxia, both in arterioles and venules. The mean difference in arterioles is 5% in terms of %SO₂ (P = 0.0027, 95% confidence interval [CI] 2%–8%, pooled data, n = 16, paired t-test). The mean difference for venules is 24% (P < 0.0001, 95% CI: 16%–32%, n = 16). 

Pseudocolor Presentation

When calibrated, the automatic software can display the results as a pseudocolor fundus image as shown in Figure 2 . Different colors represent different levels of oxygenation in retinal vessels. 

Discussion

The automatic software yields better reproducibility than does the older manual software. The system is sensitive to changes in hemoglobin oxygen saturation in both arterioles and venules when oxygen concentration in the inhaled air is changed. Furthermore, the system is easy to use and can display results as %SO₂, in numeric or graphic format. 

Table 1shows that the standard deviation for repeated measurements is much lower when the automatic software is used but the average ODR values are similar between manual and automatic analysis. When using the automatic software, the standard deviation (in terms of %SO₂) for repeated measurements is 3.7% and 5.3% for arterioles and venules, respectively. The mean saturation for this dataset was 99% in arterioles and 52% in venules. Comparing these means with prior results from researchers using other retinal oximeters is of limited value, because we used these prior means to calibrate our oximeter (see the Results section). The reproducibility can, however, be compared with prior results, although this comparison is somewhat limited by the different methodologies applied. Delori ³ used an oximeter that scanned a small region of the fundus with three wavelengths of light. The mean standard deviation between two consecutive measurements of the same vessel was 2.1% SO₂, similar for arterioles and venules (means, 98% and 45%, respectively). As 1.4 seconds were needed to achieve a single measurement, a vessel-tracking system was needed to compensate for eye movement. Using an oximeter based on scanning laser ophthalmoscopy, Smith et al. ⁶⁰ reported that in preliminary measurements on humans at a single retinal position, the measurement repeatability was typically within ±5%SO₂ (venular SO₂, typically 45%–55%). The same oximeter is reported to complete a scan in one-thirtieth of a second, ⁶¹ and a typical scan appears to involve two adjacent vessels. ⁶⁰ Schweitzer et al. ⁵⁸ used an oximeter that scans a 1.5-mm by 40-μm slit of the fundus (one or two adjacent vessels) and measures the reflectance from the vessel at wavelengths from 450 to 700 nm in 2-nm increments. Nineteen measurements on each vessel gave an SD of 4.6% SO₂ for arterioles and 4% SO₂ for venules (means, 92% and 58%, respectively). Therefore, the reproducibility of our oximeter is comparable to that in earlier reports. 

The results also demonstrate the sensitivity of the oximeter. The measured arteriolar hemoglobin oxygen saturation follows changes in inspired oxygen concentration. In hyperoxia the increase in hemoglobin saturation is limited by the fact that hemoglobin is almost 100% saturated in normoxia. In hyperoxia, oxygen that is dissolved in choroidal blood diffuses into the retina in larger amounts ^{22 62 63 64} and reduces the extraction of oxygen from the retinal circulation, thus elevating the venular oxygen saturation, as our results show. 

The retinal oximeters mentioned herein have also been shown to be sensitive to changes in oxygen saturation. This sensitivity is essential, of course, if an oximeter is to be of value for measuring changes associated with disease or treatment. Whether the oximeters show the true oxygen saturation is another matter that is difficult to resolve because absolute values are very difficult to obtain in humans. Measurements on several mammalian species with vascular retinas revealed that oxygen tension (Po ₂) of the retina and optic nerve is around 20 mm Hg when this is measured with an electrode at the vitreal surface of the retina and optic disc. ^{7 65 66 67} Assuming normal blood parameters, our venular SO₂ converts to Po ₂ of approximately 29 mm Hg. ⁶⁸ As could be expected, our venular Po ₂ value is a little higher than the tissue Po ₂ from the animal experiments (the Po ₂ gradient should be toward the tissue and away from the arterioles, capillaries, and probably venules). 

We have room to improve both sensitivity and reproducibility, because we have not yet incorporated correction for confounding factors such as light-scattering of blood cells, ³ different light paths through vessels, ⁶⁹ vessel width, ² and fundus pigmentation, ^{2 60} all of which have been shown to affect retinal oximetry to some extent. Some of the corrections require more than two wavelengths of light. Our system already has the capability to use four wavelengths simultaneously although the results presented were generated with only two. We have found better reproducibility with data based on 586 and 605 nm, compared with those based on 542 and 558 nm. 

In summary, we have shown that the automatic software improves the reproducibility of our oximeter and that the oximeter is sensitive to changes in hemoglobin oxygenation. The advantages of our retinal oximeter are that the measurement area is large, and little expertise is needed to acquire a pseudocolor fundus image of retinal vessel oxygenation. We hope this device will be useful in the evaluation of ischemic retinal diseases and their treatment. 

 Supported by Icelandic Research Council Grants 050409011 and 051220005, The Landspítali—University Hospital Fund for Science, The Research Fund of The University of Iceland, and The Icelandic Fund for Prevention of Blindness.

 Submitted for publication January 13, 2006; revised May 9 and July 19, 2006; accepted September 18, 2006.

 Disclosure: S.H. Hardarson, Oxymap (E, P); A. Harris, None; R.A. Karlsson, Oxymap (E, P); G.H. Halldorsson, Oxymap (E, P); L. Kagemann, None; E. Rechtman, None; G.M. Zoega, None; T. Eysteinsson, Oxymap (I, P); J.A. Benediktsson, Oxymap (I, P); A. Thorsteinsson, None; P.K. Jensen, None; J. Beach, Oxymap (I, P); E. Stefánsson, Oxymap (I, P)

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

 

Figure 1.

 

(A) The retinal oximeter. A traditional fundus camera, coupled with a beam splitter and a digital camera is used to obtain four images simultaneously. (B) Each of the four images is obtained at a different wavelength of light.

Table 1.

 

The Mean and Standard Deviation of Five Images of the Same Individual, Expressed in ODR Values

	Arterioles			Venules
		Manual Measurement	Automatic Measurement	Manual Measurement	Automatic Measurement
Mean	0.178 (0.065–0.300)	0.185 (0.123–0.231)	0.478 (0.360–0.640)	0.515 (0.401–0.624)
SD	0.060 (0.031–0.098)	0.026 (0.012–0.047)	0.057 (0.018–0.097)	0.037 (0.014–0.052)

 Ten individuals, one arteriole and one venule measured in each. Mean and standard deviation was calculated separately for each individual (from five images). The table shows means and range for these individual means and standard deviations.

Table 2.

 

The Mean and Standard Deviation of Five Images of the Same Individual, Expressed in SO₂ Values^*

	Arterioles	Venules
Mean	99% (93%–108%)	52% (37%–68%)
SD	3.7% (1.8%–6.7%)	5.3% (2.0%–7.3%)

 Data are as described in Table 1 .

*  Calculated by SO₂ = 125–142 · ODR.

Table 3.

 

Calculated %SO₂

	Arterioles			Venules
		Normoxia	Hyperoxia	Normoxia	Hyperoxia
Reykjavik (n = 4)	95 ± 2	99 ± 3	52 ± 2	76 ± 8
Indiana (n = 12)	96 ± 10	101 ± 9	55 ± 16	79 ± 17
Pooled (n = 16)	96 ± 9	101 ± 8	55 ± 14	78 ± 15

 Healthy subjects inhaled either room air (normoxia) or 100 O₂ (hyperoxia). Data are the mean percent ± SD.

Figure 2.

 

A screen image produced by the automatic software from a diabetic patient with no retinopathy. The color code indicates the oxygenation in the vessels.