Abstract
Purpose.:
Studies have used central retinal arteriolar (CRAE) and central retinal venular (CRVE) calibers, measured from images produced with computerized image analysis, to detect risk factors for systemic diseases. The authors explored suboptimal image focus as a possible contributing factor to artificially larger vascular caliber measurements.
Methods.:
From the reading center image collections, 30 digital retinal images were selected for optimum quality. Image analysis software was used to derive nine progressively blurred versions of the originals. IVAN measurement software was used to measure CRAE and CRVE in the original and the blurred series derived from them. To check the adequacy of the simulation, progressively defocused series of images were taken of several volunteers.
Results.:
For CRAE, each level of simulated blurring produced a statically significant increase in apparent vessel caliber from the original (P < 0.01, Wilcoxon signed rank test). For an average CRAE of 160 μm, mean broadening with minimal/moderate/severe blurring was 3 μm/12 μm/33 μm. For CRVE, every blurring level beyond the first was found to be significant (P < 0.01). From an average CRVE of 200 μm, mean broadening ranged from 0 to 11 μm with minimal to severe blurring. Analysis of the progressively defocused series taken of volunteers yielded similar results overall.
Conclusions.:
Suboptimal focus can result in erroneously larger vessel caliber measurements. Slight blurring has a minimal effect, but more severe blurring has a progressively greater effect.
The Atherosclerosis Risk in Communities Study (ARIC) introduced a computerized method of measuring and summarizing retinal vascular caliber from fundus photographs.
1 Several large population-based studies, including ARIC, the Beaver Dam Eye Study,
2,3 the Blue Mountain Eye study,
4 and the Wisconsin Epidemiologic Study of Diabetic Retinopathy,
5 have found abnormal retinal vascular caliber (either broadened or narrowed) to be associated with, or even a risk factor for, various systemic disease outcomes. Although the measurement procedure itself has been shown to be reproducible,
4 the impact of photograph quality, in particular the sharpness of camera focus, has not been sufficiently explored.
We tested this question with IVAN, a semiautomated program for measuring retinal vessel diameter created at the University of Wisconsin at Madison
5 –7 and now widely used by other investigators. In brief, an algorithm for segmenting vessels from the retinal pigment epithelium (RPE) divides a color image into separate homogeneous regions called splats. Afterward, a vessel-tracking module determines the margins of presumed vessels based on the splat boundaries and measures their widths. (Of note: though color information is used to type vessels as to arterioles or venules, vessel caliber is determined from only the monochromatic green channel, which offers highest contrast of vessels against RPE.)
By optical principle, defocusing attenuates the highest frequencies in an image, artificially widening the apparent dimensions of features in that image compared with an optimally focused image.
8 Briefly, defocusing entails a point spread function in which some light rays reflected from the object of interest appear to the observer to be originating outside the actual margin of that object, in effect broadening its apparent dimensions. Thus, if a fundus photograph is taken with less than optimal focus, we might predict that the apparent vessel caliber would be artificially widened. In the case of the vessel segmentation approach used in IVAN, we might expect to see artificial expansion of the splats corresponding to vessels.
Using well-focused images with established vascular measurement results, we used image editing software (Photoshop CS3; Adobe, San Jose, CA) to model progressive defocusing of the image. With the derived series of images, we measured the impact of defocusing on the retinal vascular caliber results. To check the adequacy of the simulated blurring, we also took a series of progressively defocused fundus photographs of several volunteers by graduated movement of the focus knob from the optimal setting.
A sample of 30 native digital TIFF images was selected from various digital image databases at the Fundus Photograph Reading Center (FPRC). These databases were compiled during multicenter research studies, all conducted in accordance with the tenets of the Declaration of Helsinki and with institutional review board approval. All patient identifying information had been made anonymous at study clinical centers before it was sent to the FPRC, where the images were indexed only by coded study identifiers.
Images were selected randomly but had to meet criteria for excellent photographic quality based on the subjective assessment of both an FPRC image evaluator and an experienced FPRC consultant photographer. The chosen eyes had clear media (i.e., no detectable lens opacity) and minimal or no abnormality that would interfere with vascular measurement. Where a stereo pair was present, the image with the better photographic quality was chosen. This was subjectively determined based on the following criteria: focus (perceived sharpness of the vessel edge against the background retinal pigment epithelium), illumination (even and consistent lighting across the zone of vessel measurement), color balance (equal saturation of color channels).
The selection included images acquired by various digital camera systems from multiple clinical sites, across a number of disease types. The FPRC requires that digital cameras have at least 3 megapixels (MP) of spatial resolution; most have 5 MP or greater. All digital camera systems and photographers were certified by the FPRC to ensure that they complied with standardized protocols.
Image editing software (Photoshop CS3; Adobe) was used to systematically degrade image focus of the original images by use of the box blur filter. The box blur filter was determined to have more usefulness than the regular blur filter because of its ability to scale the blurring effect along a linear gradient. Radius of blurring ranged from one through nine adjoining pixels. We chose nine as the upper limit of the pixel radius for the blurring effect because anything beyond that would have been clearly poor quality for evaluation. After the desired level of blurring was applied, the image was saved as a new TIFF file. The original file was then loaded again, and the next step of degradation was completed until we had a series consisting of the original image and nine progressively blurred variants for each subject.
A certified and experienced grader used IVAN (version 1.3) to measure the central retinal arteriolar (CRAE) and the central retinal venular (CRVE) calibers of the original images. The grading procedures have been described in detail elsewhere.
5 A screen shot of the baseline measurements were used as a guide to ensure that grid placement remained consistent and that vessels were measured in similar segments. This was done to reduce the natural variability of vessel caliber over the course of the vessel on our measurements.
A Wilcoxon signed rank test was performed comparing the CRAE and CRVE of each blur level to the original (
Table 1). Because of the small sample size of the study, we could not assume that the population was normally distributed. As such, a nonparametric test was deemed superior.
Table 1. Wilcoxon Signed Rank Test of Blurred Derivations versus Original Image
Table 1. Wilcoxon Signed Rank Test of Blurred Derivations versus Original Image
| CRAE P | CRVE P |
| Blur 1 vs. original | <0.01 | 0.81 |
| Blur 2 vs. original | <0.01 | <0.01 |
| Blur 3 vs. original | <0.01 | <0.01 |
| Blur 4 vs. original | <0.01 | <0.01 |
| Blur 5 vs. original | <0.01 | <0.01 |
| Blur 6 vs. original | <0.01 | <0.01 |
| Blur 7 vs. original | <0.01 | <0.01 |
| Blur 8 vs. original | <0.01 | <0.01 |
| Blur 9 vs. original | <0.01 | <0.01 |
To test the theory of blurred images (Photoshop; Adobe) behaving similarly to those that are naturally defocused, an internal proof-of-concept exercise was also performed.
Three volunteers had a series of retinal photographs of both eyes taken by an experienced ophthalmic photographer. Using a digital fundus camera (50 DX; Topcon, Yamagata, Japan) with a dial focus knob, we created an indicator for the focus knob and used a protractor to consistently increment the focus intervals. After establishing a baseline photograph of excellent focus quality, as subjectively determined by our photographer, the focus knob was moved in 3° intervals in a clockwise direction for five successive turns. The resultant photographs were then similarly measured for retinal vascular caliber using IVAN software.
Our results demonstrate that variability in camera focus by the photographer has a measurable effect on apparent vessel calibers. Slight blurring has minimal effect, but more severe blurring (but still within the range of quality typically received by the FPRC) has progressively greater effect. This artificial broadening from defocusing is particularly apparent with CRAE, although CRVE is also significantly affected.
Results from published epidemiologic studies can be used as a frame of reference for assessing the impact of the broadening effect from defocusing of images. In this instance, data from The Multi-Ethnic Study of Atherosclerosis are used, shown in the Appendix.
9 The significant interquartile ranges between groups are consistently 1 to 2 μm. When taken in the context of this study, it is obvious that a shift of even a few micrometers in vessel caliber caused by defocusing could result in subjects with defocused images being classified into the wrong quartile. This is representative of the sensitive nature of vascular measurement and the possible impact that even a slight systematic shift in micrometers could have on results.
The CRAE was shown to be more vulnerable than the CRVE to a focus effect. One possible factor in this difference of effect is the lower contrast of arteriolar blood against RPE background compared with that of venular blood. Relative vessel size may also contribute. Arterioles are typically smaller than venules, and defocusing may have a relatively greater effect as vessel size vessel decreases.
One might expect that in a large population-based study, the effects of focus would be equally distributed among groups because of the large sample size. However, this assumes that focus affects all groups equally. If some groups are disproportionately affected by focus, then comparisons of those groups would less accurate. For example, others
9,10 have reported that there may be a pigmentation effect that systematically, and perhaps artificially, impacts vascular caliber measurement. Our study was not designed to investigate the effects of focus relative to retinal background pigmentation. However, as we have demonstrated experimentally, defocusing of images within the normal range of image quality could account for an effect of the magnitude reported. One might expect that lower contrast of the dark (blood-filled) vessels against greater RPE pigmentation might make it more difficult for photographers to achieve sharp focus on the vessels consistently. Furthermore, our observation of the overall greater impact from defocusing on CRAE (lower RPE contrast) compared with CRVE (higher RPE contrast) suggests that the defocusing effect might be exacerbated when both CRAE and CRVE contrast against an RPE background is reduced. Further investigation is warranted to illuminate this topic.
In large epidemiologic studies, vascular measurement has already been shown to be useful for identifying associations with, and even prediction of, important systemic disorders. However, in such studies, it is difficult to isolate and study a particular methodological effect systematically. We have been able to approach this question through a simulation study by looking at the magnitude and thresholds of the impact of focus on measured vascular caliber. Our results suggest that using vascular measurement, as currently performed, to make close comparison of racial groups may be problematic.
Provided a study includes enough subjects to be reasonably powered, one might assume that the subgroups of interest would be equally affected by measurement “noise” (e.g., artifactually broader measurements). Consequently, the differences between groups in mean CRAE or CRVE would still be real, although the absolute values might be slightly inflated. In contrast, however, if there are systematic differences between subgroups in factors that affect sharpness of focus (e.g., degree of pigmentation or lens opacity), the apparent differences between groups in vessel caliber may have to be examined further.
Based on our findings, studies might consider tightening quality criteria for what constitutes acceptable focus. Although this appears desirable, it may not be feasible. In the real-world setting of academic clinics and examination centers, not every photographer will be technically skillful enough, using currently available equipment, to obtain sharp focus for every subject. Nor might imposing stricter focus criteria be methodologically sound. There may be real differences between subgroups being compared that affect sharpness of focus. For example, the theory that the eyes of darker pigmented persons may be more challenging to focus sharply because they may not dilate as widely and because of the contrast between vessel and RPE the photographer uses for fine adjustment is less obvious.
Our proof-of-concept exercise showed that the image editing software format of artificially defocusing photographs using the box blur function has a similar impact on the ability of IVAN to accurately measure vessel width as natural photographic focus. Although we cannot draw a precise comparison between the two, because we cannot relate levels of blurring in image editing software to degrees of adjustment on a focus knob, both show obvious increases in vessel width as defocusing increases, although at somewhat different magnitudes.
One limitation of our study is the subjective determination of focus. There is no method to quantify the level to which a photograph is defocused. Ideally, we would have an objective cutoff, after which images would be considered unusable. Another limitation of our study is that we performed it on a relatively small number of subjects. However, the phenomenon we are exploring is explainable through commonly accepted optical principles. Thus, we would not expect the measurement of additional samples to change the tenor of our results. We have provided an introductory look at what could be a major factor that has yet to be widely taken into consideration when looking at results from vessel caliber measurements.
There are several possible avenues of future study in this area. Media opacities also create a blurring effect on retinal photographs. We are particularly concerned about lens changes (nuclear sclerosis, cortical and posterior subcapsular opacities). In theory, the blurring effect of media opacities is similar to that produced by a defocused camera. Thus, one would expect that vessel caliber would appear to be wider in eyes with greater media opacity. This could complicate subgroup comparisons for which systematic differences in media opacity are likely. We have yet to develop an effective method for simulating the effect of media opacity, but the possibility remains intriguing.
To address the problem of the subjectivity of observer assessment of focus, we are attempting to develop a more objective and quantitative metric. For example, we have been looking at adapting the full-width half-height parameter used to evaluate focus in astronomical images. Ultimately, we hope to model the impact of different degrees of measurable defocusing so that we can develop and apply a method for post hoc correction of measurements on less than optimal images.
The development of a real-time focus metric would help to objectively determine the quality of a photograph during the session when it could be retaken. This would eliminate the need for subjective photographer judgment. An alternative would be to create a procedure that would allow us to quantify degree of defocusing in existing images and to estimate the actual dimensions of vessels based on a pragmatic correction model. A correction model would allow us to not only to determine quantifiable acceptability regarding focus but also to salvage unacceptable photographs by applying post hoc correction. This would also help us to minimize any differences caused by systematic defocusing among groups.
In conclusion, defocusing of retinal images artificially broadens the apparent diameters of both arterioles and venules, conceivably affecting outcomes studied in ocular epidemiology and clinical research. Investigators may have to consider differential image quality as a technical factor that could systematically impact between-group comparisons in vascular measurement studies.
Disclosure:
C.S. Chandler, None;
S. Gangaputra, None;
L.D. Hubbard, None;
N.J. Ferrier, None;
T.W. Pauli, None;
Q. Peng, None;
D.W. Thayer, None;
R.P. Danis Jr., None