IOVS
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Originally published In Press as doi:10.1167/iovs.07-1267 on April 17, 2008
 (Investigative Ophthalmology and Visual Science. 2008;49:3269-3282.)
© 2008 by The Association for Research in Vision and Ophthalmology, Inc.
doi:10.1167/iovs.07-1267
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
iovs.07-1267v1
49/8/3269    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Web of Science (2)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Hubbard, L. D.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Hubbard, L. D.

Brightness, Contrast, and Color Balance of Digital versus Film Retinal Images in the Age-Related Eye Disease Study 2

Larry D. Hubbard,1 Ronald P. Danis,1 Michael W. Neider,1 Dennis W. Thayer,1 Hugh D. Wabers,1 James K. White,1 Anthony J. Pugliese,2 Michael F. Pugliese2 for the Age-Related Eye Disease 2 Research Group

1From the Department of Ophthalmology and Visual Sciences, University of Wisconsin-Madison, Madison, Wisconsin; and 2Choices for Service in Imaging, Inc, Glen Rock, New Jersey.


   Abstract
Top
 Abstract
Population and Methods
 Results
 Discussion
 Conclusions
 References
 
PURPOSE. To analyze brightness, contrast, and color balance of digital versus film retinal images in a multicenter clinical trial, to propose a model image from exemplars, and to optimize both image types for evaluation of age-related macular degeneration (AMD).

METHODS. The Age-Related Eye Disease Study 2 (AREDS2) is enrolling subjects from 90 clinics, with three quarters of them using digital and one quarter using film cameras. Image brightness (B), contrast (C), and color balance (CB) were measured with three-color luminance histograms. First, the exemplars (film and digital) from expert groups were analyzed, and an AMD-oriented model was constructed. Second, the impact of B/C/CB on the appearance of typical AMD abnormalities was analyzed. Third, B/C/CB in AREDS2 images were compared between film (156 eyes) and digital (605 eyes), and against the model. Fourth, suboptimal images were enhanced by adjusting B/C/CB to bring them into accord with model parameters.

RESULTS. Exemplar images had similar brightness, contrast, and color balance, supporting an image model. Varying a specimen image through a wide range of B/C/CB revealed greatest contrast of drusen and pigment abnormalities against normal retinal pigment epithelium with the model parameters. AREDS2 digital images were more variable than film, with lower correspondence to our model. Ten percent of digital were too dim and 19% too bright (oversaturated), versus 1% and 4% of film, respectively. On average, digital had lower green channel contrast (giving less retinal detail) than film. Overly red color balance (weaker green) was observed in 23% of digital versus 8% of film. About half of digital (but fewer film) images required enhancement before AMD grading. After optimization of both image types, AREDS2 image quality was judged as good as that in AREDS (all film).

CONCLUSIONS. A histogram-based model, derived from exemplars, provides a pragmatic guide for image analysis and enhancement. In AREDS2, the best digital images matched the best film. Overall, however, digital provided lower contrast of retinal detail. Digital images taken with higher G-to-R ratio showed better brightness and contrast management. Optimization of images in the multicenter study helps standardize documentation of AMD (ClinicalTrials.gov NCT00345176).

The Age-Related Eye Diseases Study 2 (AREDS2), a multicenter clinical trial of nutritional supplements to inhibit advanced AMD, uses both film and digital cameras to image the retina. Methodological continuity with the predecessor AREDS, which used only film cameras, is crucial so that data from the two studies are comparable. Specifically, the technical parameters of digital images must be adequate to allow the sensitivity and reproducibility of AMD evaluation observed with film photographs in AREDS (reports 6 and 171 2 ). In published reports from film/digital comparison studies at single centers, there is good agreement on grading of AMD presence and severity between film and digital images of the same subjects within one clinic.3 4 5 However, reading centers for multicenter studies encounter different clinics having a wide range of equipment and technicians. Thus, it is necessary to analyze the characteristics of digital images versus film photographs so that they can be standardized as much as possible.

This report analyzes the parameters of digital versus film images in the AREDS2, and compares their suitability for AMD evaluation. In particular, we focus on basic aspects of image quality: brightness (B), contrast (C), and color balance (CB). By "brightness," we mean the degree of image exposure (from under to over), by "contrast" the degree of apparent difference between retinal features and background, and by "color balance" the relative strength of the constituent color channels.

Our goal was to develop a framework to compare the B/C/CB parameters of digital versus film color fundus images for subjective evaluation of AMD abnormalities, and to manage these parameters to obtain optimal standardized images. Our approach is based on objective measures of B/C/CB parameters by using the three-color luminance histogram. First, we analyzed exemplar images selected from a variety of sources, both film and digital, and formulated an image model oriented toward AMD detection. Second, we systematically varied the B/C/CB parameters in an AREDS2 image with obvious AMD content to demonstrate the effect on disease appearance. Third, we compared B/C/CB between AREDS2 film and digital images, determining their range and distribution. Fourth, we developed a histogram-based enhancement procedure for suboptimal AREDS2 images, both natively digital and digitized film, to optimize them for AMD grading.

Spatial resolution is also an important aspect of image quality and in the past was of concern when digital cameras produced fewer megapixels. However, because ophthalmic cameras with high-resolution sensors are commonplace in ophthalmic clinics, our study satisfied this need by mandating digital cameras with resolution near to or equal that of film. For documentation of AMD (with drusen as small as 32 µm diameter), imaged at the 30° or equivalent magnification setting, 3-megapixel systems are acceptable, with ≥6 megapixels preferred. Such sensors yield resolving power on the retina of approximately 20 and 13 µm, respectively, by pragmatic calculation.

For film imaging, clinical trials achieved consistent B/C/CB by specifying the acceptable film emulsions and development processes. Digital images can be more variable than film, depending on the model of the camera, the capture software settings, and how the photographer adjusts these setting at the time of photography. This report formulates and explores an approach for analyzing and managing B/C/CB of digital fundus images to preserve the consistency of image documentation historically achieved by film, and to improve this capability when possible by taking advantage of the potentially enhanced detection of lesions with optimized digital images. We developed a novel optimization procedure based largely on normal image content because existing techniques, such as auto-optimization, can be overly sensitive to the disease content in retinal images, and thus produce highly variable results between subjects or within the same subject over time.


   Population and Methods
Top
 Abstract
 Population and Methods
Results
 Discussion
 Conclusions
 References
 
AREDS2 (currently under way) plans to recruit and observe 4000 subjects. Eligibility characteristics include: (1) both males and females of any race; (2) 50 to 85 years of age at entry; (3) bilateral large drusen (≥125 µm diameter), or large drusen in one eye and advanced AMD (either choroidal neovascularization or central geographic atrophy) in the fellow eye; and (4) fundus photographs of adequate quality for confident evaluation of AMD status (dilation of ≥5 mm, sufficiently clear media, and ability to cooperate with photography). Study subjects gave written informed consent, all participating institutions received institutional review board approval, and the study is being conducted in accordance with HIPAA (Health Insurance Portability and Accountability Act) requirements and the tenets of the Declaration of Helsinki.

Retinal Imaging Equipment and Procedures
Subjects were photographed at 90 clinical centers, including academic institutions and private practices. Fundus photographers were required to become formally certified by submitting satisfactory specimen photographs taken on nonstudy subjects. Camera systems in the various clinical centers include both film and digital media, in an approximately 1:3 ratio. For film photography, clinics are required to use cameras approved by the Reading Center (models FF3-4 and FF450; Carl Zeiss Meditec, Oberkochen, Germany; and Topcon 50-XT; Topcon Medical Imaging Corp., Tokyo, Japan). (We approved other makes of cameras, but they were not used in this study.) Approved film emulsions and development were mandated (Ektachrome 100 Professional slide film or equivalent, developed by Kodak-certified Q-Laboratories; Eastman Kodak, Rochester NY). Digital clinics were required to use Reading Center–approved cameras checked via equipment certification (Visupac, Carl Zeiss Meditec; Topcon IMAGEnet, OIS [Ophthalmic Imaging Systems], Sacramento, CA; Escalon/MRP, Escalon Medical Corp., Wayne, PA; and DHC [Digital Health Care], Cambridge, UK). All these incorporate area-array, silicon-based, charge-coupled device (CCD) Bayer sensors (single chips with pixels individually filtered to detect red, green, or blue [R/G/B] in a 1:2:1 ratio) of various makes. (Although three-chip sensors, containing separate sensors for red, green, and blue, with sufficient spatial resolution would have been acceptable, no AREDS2 clinics used them for the study.) Photographers are allowed to use their customary settings for B/C/CB, provided the Reading Center has approved the appearance of samples. The choice of imaging with film versus digital was made by the clinic based on equipment availability and local preference.

The imaging procedure has been described previously in AREDS Report 6.1 Briefly, it includes stereoscopic 30° (or equivalent) color fundus photographs taken through a pharmacologically dilated pupil: field 1M (disc), field 2 (macula), field 3M (temporal to macula), and fundus reflex (anterior segment).

Handling and Display of Images
The Reading Center is digitizing all AREDS2 film transparency slides, to integrate the flow of images within an all-digital environment and to allow the use of digital tools (e.g., planimetric measurement of the areas of various abnormalities). A previous report by Scholl et al.6 indicated that grading digitized film images for AMD produced similar results to grading the original film. The procedure for digitization was carefully developed by using the film standard photographs from AREDS and sample film photographs from AREDS2. The scanning parameters have been set so that the overall tonal appearance on the digital monitor closely matches that of the original slide viewed on the standard Reading Center light box (containing three 14-watt daylight fluorescent tubes; color temperature, 6300°K) from the original AREDS. Our standard digital monitor is an LCD 20.5-in. display, set with gamma of 2.2, color temperature of 6500°K, and luminance of 125 candelas, all checked monthly with an external calibration system (Greytag Macbeth; X-Rite Incorporated, Grand Rapids, MI). Slides are digitized (Super CoolScan model 5000 ED; Nikon Corp, Tokyo, Japan), with autofocus enabled at 3400 x 2300-pixel resolution.

Digital images are displayed (IMAGEnet system, ver. 2.56; Topcon) and customized with additional tools to analyze and manage red/green/blue (RGB) parameters: a three-color channel histogram display, with sliders to adjust the brightness and contrast of the individual color channels before recombining them into a modified image. (The original image is retained, and the grader can compare it to the enhanced image.) All 24-bit RGB images are sent from the clinic in TIFF format (uncompressed), but are stored at the Reading Center in 2912 x 2480 format with maximum-quality JPEG compression (approximately 20:1) to conserve file space. (Lee et al.7 reported that careful comparison of color images in TIFF [uncompressed] and JPEG [compressed 30:1, which they considered "low compression"] format were "virtually indistinguishable" as to spatial and color resolution, stereo effect, and drusen grading result, whether performed manually as in AREDS2 or automatically with a segmentation program.) Images from other makes of digital systems are transferred into IMAGEnet as a single uniform environment for display and analysis. For the figures in this report, images were analyzed and modified with commercial software (PhotoShop CS2; Adobe Systems, San Jose, CA).

Image Analysis for B/C/CB Parameters
Digital color images were analyzed by using three-color channel luminance histograms (Fig. 1) to analyze and manipulate separately each of the RGB channels. Each channel is represented by a luminance curve (distribution of intensity or brightness) on a 256-step scale (from 0 or the "black point," at which the sensor detects no light, to 255 or the "white point," at which the sensor is saturated by light), constituting the camera’s dynamic range. In conventional luminance curves, the x-axis shows intensity values from 0 to 255, and the y-axis shows the number of pixels in the image with a given intensity value. RGB information composes histogram curves, which are approximately bell shaped, and exhibit a peak, a main mass, and left and right tails. To facilitate description and analysis, the 256-intensity scale has been summarized into 16 steps, each containing 16 intensity levels.


Figure 1
View larger version (73K):
[in this window]
[in a new window]

  		FIGURE 1. AREDS standard photograph 12 (left side of stereo pair), with luminance histogram of its three composite color channels (RGB), and with monochrome of the three channels to show their relative brightness and contrast. (1) This image utilizes most of the dynamic range, with the composite curves extending from approximately 1/16 on the dark end of the dynamic range to 15/16 on the bright end. (2) The green and red channels contribute most of the contrast information, each curve spanning approximately 8/16 of the dynamic range, measured tail to tail. (3) The green and red channels are offset from each other so that they bracket the middle of the dynamic range, green occupying mostly the lower half (from 2/16 to 11/16, peaking at 6/16) and red mostly the upper half (from 7/16 to 15/16, peaking at 12/16), with only slight overlap between them at the midpoint. (4) The G/R color balance is approximately 0.5, defined as the ratio of the locations of the G and R curve peaks in the dynamic range. (5) The blue curve is at the lower end of the dynamic range and has a narrow span (from 9/16 to 4/16—its mass concentrated between 1/16 and 2/16), with a B/R color ratio of 0.20 (defined similarly to G/R ratio). Reprinted with permission of the University of Wisconsin-Madison Fundus Photograph Reading Center on behalf of the Age-Related Eye Disease Study (AREDS) Research Group.

 
We define B/C/CB parameters in terms of this collapsed 16-step system. For each color channel, "brightness" is defined as the location of the luminance curve peak in the range (essentially its mode), and "contrast" as the span of the curve on the range. For example, the red channel curve in Figure 1 peaks around 12/16, indicating its brightness, and has a span of approximately 8/16 (from the left tail at 7/16 to the right tail at 15/16, in effect its range), signifying its contrast. In this simplified system, G/R color balance is defined as the ratio of the green curve peak location to the red curve peak location (i.e., green brightness compared with red brightness). Thus, an image with a green curve peak at 6/16 and a red curve peak at 12/16 has a G/R ratio of 0.50 (6/12). To account for the blue component, the B/R ratio is similarly defined as the ratio of blue curve peak location to red curve peak location. (To reduce rounding error in determining color balance ratio, which became apparent with the 16-step scale due to broadness of each step, we instead used 32 steps of 8 levels each when calculating this parameter.)

To compute luminance histogram characteristics across large samples of images, we used custom image analysis software contracted from Choices for Service in Imaging, Inc. (Glen Rock, NJ). This processor reads the pixels in vertical columns (every fourth column, for efficiency) and assigns the RGB intensity values of each pixel into one of 32 luminance categories. The processor detected the location of the fundus image within the frame (variable across digital and film cameras) and established a region of interest (ROI) excluding not only the black frame but also an annulus 1.8 mm across (one standard disc diameter) from the periphery of the fundus image. (This trimming minimized edge artifacts, much as hardware cones do in film cameras, and typically excluded the disc, which is much brighter than the macula, from the macular image.) Finally, the processor summarized the peak location and span of each of the color curves for the image of interest and then created a thumbnail of the image for reference. Results for all images in the sample were compiled into a spreadsheet (Excel; Microsoft Inc., Redmond, WA) for querying and analysis.

To summarize B/C/CB across the exemplar images, we extracted luminance histograms, calculated the parameters as defined above, and determined the mean, the median, and the range of their values. Using this image "recipe" as a starting point, we further adjusted the parameters by inspection so that the final AMD model appeared to represent the best features of the various exemplars.

To simulate the effect of these different B/C/CB parameters on AMD appearance, we selected a specimen from the AREDS2 sample showing typical AMD abnormalities: drusen, increased RPE pigment, and depigmentation. Using the Photoshop Exposure tool (from the Image/Adjustments menu), we were able to mimic different degrees of exposure during the imaging session, thereby controlling brightness and contrast. With the Brightness/Contrast tool, we were able to shift the color curves in relation to each other, thus controlling color balance. The modifications were driven by luminance histograms, so that the resultant array of simulated images matched the brightness (with associated contrast) and color balance observed in the actual sample of AREDS2 digital images. We measured the contrast of AMD abnormalities to adjacent normal retinal pigment epithelium (RPE), calculating the absolute difference in mean luminance levels between 3 x 3-pixel areas sampled at the center of the abnormality and just outside its boundary.

AREDS2 film and digital images were similarly analyzed via luminance histograms, so that their distributions of B/C/CB could be displayed graphically and expressed statistically (median, range, and conventional percentiles). We constructed an illustrative array of film and digital images, with the rows showing different G/R color balance ratios and the columns showing the percentiles of brightness. (Because the red channel is brightest, its curve peak location was taken to represent the image.)

Standardized Enhancement of Retinal Images
Substandard AREDS2 images were enhanced ("optimized") by extracting their luminance histograms, then using the custom IMAGEnet tool Histogram RGB Channels (similar to the Levels tool in the Photoshop Image/Adjustments menu) to manipulate the individual color channel curves until they accorded as closely as possible to the those of our image model.

Statistical Analysis
Differences in distribution of B/C/CB parameters between study digital and film images were tested for significance with the Wilcoxon test for location (SAS Institute, Cary, NC).


   Results
Top
 Abstract
 Population and Methods
 Results
Discussion
 Conclusions
 References
 
For all figures displaying fundus images, the reader is encouraged to download the PDF version of this article from the IOVS website (http://www.iovs.org) in order to view them in digital form. Recommended monitor settings are color temperature = 6500°K and gamma = 2.2 (Windows systems) or 1.8 (Macintosh systems).

Exemplars of High-Quality Fundus Images Analyzed
Selected exemplar images from three sources are displayed in Figure 2 . From the film standard photographs (SP) in the AREDS AMD Classification, we selected four images (SP4, -11, -12, and -13) because of their content and quality (Figs. 2A 2B 2C 2D) . For example, SP 12 (also seen in Fig. 1 ) depicts both of the main abnormalities characteristic of nonadvanced AMD: drusen and pigment abnormalities. (Appreciated in stereo, it also shows advanced AMD: an obvious dome of serous sensory retinal detachment, with a subtle change in retinal transparency seen as a color shift.)


Figure 2
View larger version (97K):
[in this window]
[in a new window]

  		FIGURE 2. (A–D) Four selected AREDS standard photographs (4, 11, 12, and 13, respectively), digitized from the film originals. Reprinted with permission of the University of Wisconsin-Madison Fundus Photograph Reading Center on behalf of the Age-Related Eye Disease Study (AREDS) Research Group. (E) The "best of show" award-winning color fundus photograph from the Ophthalmic Photographers Society’s (OPS) 2006 competition (retinal montage courtesy of Richard Hackel, CRA, University of Michigan-Ann Arbor, and the Journal of Ophthalmic Photography). (F–I) Four color fundus images chosen by four different major manufacturers to advertise their cameras on their Web sites. (Permission was obtained from all vendors. To maintain vendor neutrality, identifying image marks have been masked.) All images are accompanied by their three-color histograms, interpretable as described in Figure 1 . Enhanced versions according to the iMD Chrome model are shown beneath the originals.

 
To allow comparison with natively digital images from more contemporary sources, we obtained the "best of show" award winner from the 2006 competition of the Ophthalmic Photographers Society or OPS (Fig. 2E) , and advertising images from the Web sites of four major digital camera manufacturers masked as to brand (Figs. 2F 2G 2H 2I) . Inspection of the three-color luminance histograms of these images (Fig. 1) reveals that the parameters of these images are reasonably similar: (1) their luminance curves in composite occupy almost all the available dynamic range, avoiding only its very bottom and top; (2) the red and green channels have extensive spans, resulting in high contrast of features against background; (3) red and green channels are offset from each other in most cases, a relationship that appears to give both green and red relatively strong "voices" in the aggregate image; and (4) the blue curve is typically subordinate to both red and green, being low in the range and having relatively narrow span (the retina absorbs most blue light). Blue curve location and span are the most variable parameters across the exemplars.

Table 1 lists the B/C/CB parameters from the exemplar images shown in Figures 1 and 2 . We calculated the overall mean and median values (very similar) across the exemplars, giving equal weight to each of the three sources. (Although a single image, the OPS image was selected by their panel of experts from competition with many others.)


View this table:
[in this window]
[in a new window]

  		TABLE 1. B/C/CB Parameters of Exemplar Retinal Images, Summarized, Results Pragmatically Adjusted to Yield an Image Model

 
Construction of a Model Image
Based on three types of exemplar images (four AREDS standard photographs, the OPS best of show image, and four manufacturers’ advertising images), we developed an image model oriented toward reliable documentation of AMD. Beginning with the examplar mean and median values (Table 1) , we pragmatically adjusted our target parameters further, guided by the observed effect on depiction of AMD abnormalities, to achieve the most informative image. We restricted the curves from the top 1/16 and bottom 1/16 of the range, because human vision does not perceive detail well at the extremes. We adopted the mean or median values, with the following exceptions: (1) We increased the target span of both red and green curves to 8/16, approximately the maximum observed among the exemplars, to enhance contrast of retinal detail against RPE background. (Beyond this point, we judged that further enhancement introduced artifactual texture to the relatively homogenous RPE, confounding detection of retinal detail). (2) We suppressed the blue channel brightness and contrast, and thus the B/R color balance ratio, toward the minimum values seen in the examplars, because blue carries little retinal information. We judged that higher blue content introduces more "noise" into the image and gives it a purplish cast that muddies its appearance.

The final model, dubbed iMD Chrome, was constructed with the histogram parameters presented in the bottom row of Table 1 : red curve span from 7/16 to 15/16 (peak near 12/16), green curve span from 1/16 to 9/16 (peak around 6/16), and blue curve span from 1/16 to 3/16, (peak around 2/16), yielding a G/R ratio of 0.50 and a B/R ratio of 0.17. Figure 2 shows the results of this formula applied to the exemplar images (below the originals). Examining the histograms as a gestalt, the red and green curves appear equally broad, with the red curve occupying mostly the upper half and the green curve occupying mostly the lower half of the upper range, with slight overlap at the range midpoint. The blue curve is a narrow spike anchored to the left end of the green curve.

Effect of B/C/CB on AMD Appearance
Because AREDS2 is studying age-related macular degeneration, we tested the effect of variance in B/C/CB parameters on the appearance of typical AMD abnormalities against the normal RPE background. Using several images selected for typical AMD content, we systematically manipulated these variables (Photoshop; Adobe Systems) and then measured contrast between abnormality and background. Figure 3 presents the results for one such image (others yielded similar results). We used brightness (represented by red curve peak location) as the cardinal variable, using the gradations actually observed in AREDS2 digital images. For simplicity, we display a series with color balances fixed at the following model values: G/R ratio = 0.50 and B/R ratio = 0.17. (As expected, lower G/R ratio decreased the measured abnormality/background contrast in the green channel [not shown].)


Figure 3
View larger version (48K):
[in this window]
[in a new window]

  		FIGURE 3. Measurement of contrast of typical AMD abnormalities against RPE background in each color channel (RGB), through the AREDS2 range of illumination (10th–95th percentile, as illustrated in Fig. 5 ). (A) Three abnormalities (a druse, increased pigment, and depigmentation) selected for measurement of a 3 x 3-pixel ROI, to obtain mean RGB values. Contrast was calculated as absolute difference, in each color channel, between luminance levels of the abnormality and its adjacent background. Relative contrast of abnormality against background is plotted as a function of brightness in (B) for the druse, (C) for the increased pigment, and (D) for the depigmentation. For the image series showing this eye varied through this range of illumination, with luminance histograms, see the bottom row of Figure 6 .

 
Figure 3A marks the locations of a druse, a clump of increased pigment, and a locus of depigmentation. (Each instance was chosen to represent the midrange on the spectrum from subtle to obvious.) Within each color channel, degree of contrast was calculated as the absolute difference in brightness between the luminance values of each abnormality and those of adjacent normal RPE, as plotted in Figure 3B for the druse, Figure 3C for increased pigment, and Figure 3D for depigmentation.

For the druse, the greatest contrast was in both red and green, whereas for both increased pigmentation and depigmentation, the greatest contrast was in red with less contrast in green. (For all abnormalities, there is minimal contrast in blue.) As expected, the contrast between abnormality and background became greater as illumination increased. However, as the red channel approached oversaturation, contrast between abnormality and background decreased, somewhat for the druse, more for increased pigment, and even more for depigmentation. In fact, as more pixels reached oversaturation in red, the apparent contrast between pigment abnormalities and normal RPE background vanished (i.e., the abnormalities were obliterated). A moderately bright (but not oversaturated) image corresponding to the iMD model (between the 75th and 90th percentiles), displayed the highest contrast of AMD abnormality against normal RPE background.

B/C/CB of AREDS2 Digital and Film Images Analyzed
Samples of the first images received in AREDS2 from every clinic (the better side of the stereo pair of the right eye macula) were analyzed for brightness, contrast, and color balance via luminance histogram. Table 2 gives the distributions of B/C/CB parameters for digital (605 eyes) versus film (196 eyes). (Because subset analysis revealed that distributions of B/C/CB parameters were substantively similar across all makes of digital camera, the data were pooled across all makes.) There were highly significant differences between the two media for every image parameter tested, except for the blue component. Figure 4 plots the frequency distributions of brightness (Fig. 4A) and contrast (Fig. 4B) in each color channel (RGB) and the G/R and B/R color balance ratios (Fig. 4C) .


View this table:
[in this window]
[in a new window]

  		TABLE 2. B/C/CB Parameters of Digital versus Film Images in AREDS2

 

Figure 4
View larger version (34K):
[in this window]
[in a new window]

  		FIGURE 4. Frequency distributions of image parameters of digital (n = 605 eyes) compared with film (n = 196 eyes) color fundus images, based on the better side of the macular stereo pair in the right eye. (A) Distribution of brightness by color channel, based on location of each curve peak in the dynamic range, expressed in 16ths. (B) Distribution of contrast by color channel, based on tail-to-tail span of each curve on the dynamic range, expressed in 16ths. In both (A) and (B), colors of the plotted points are keyed to the identities of the color channels (RGB). The individual points in the plots have been connected to emphasize the profile of the distribution (solid lines for film, dashed lines for digital). Brightness and contrast distributions are significantly different by Wilcoxon test for location, highly so for red and green (P < 0.0001) and moderately so for blue (brightness, P = 0.0501; contrast, P = 0.0021). (C) Distribution of G/R and B/R color balance ratios, with darker shade for film and lighter shade for digital. Color balance ratios are defined as the color 1 peak location over the color 2 peak location, both expressed in 16ths of the dynamic range. By Wilcoxon test for location, differences between film and digital distributions for both color balance ratios are highly significant (P < 0.0001). See Figure 5 for representative images with histograms illustrating B/C/CB (chiefly overall brightness and G/R color balance) at the middle of these distributions and near their bottoms and tops.

 
Overall, B/C/CB parameters of the film images tended to be more similar to those postulated for the iMD image model (Table 1) . Comparison of image parameters between digital and film revealed both similarities and differences. Regarding brightness, both red and green median peak locations were somewhat lower in digital (at 10/16 and 5/16, vs. 13/16 and 7/16, respectively), with much greater variation of red channel brightness. Median red span (contrast) was wider in digital (6/16), but median green span was narrower (4/16). More muted green in digital resulted in a shifted distribution of G/R ratio. Although the median was only slightly lower than that on film (0.47 vs. 0.50), 23% of digital images were at 0.35 or lower, versus 9% of film images (data not shown). Less muted blue in digital resulted in a shifted distribution of B/R ratio, with a median of 0.21 versus 0.16 for film.

When we examined B/C/CB parameters within individual clinics (data not shown), we found that, unlike film clinics and those digital clinics with consistently optimal images (the minority), digital clinics with fewer optimal images (the majority) revealed marked intraclinic variability in B/C/CB parameters (e.g., some images underexposed, others overexposed). We typically could not discern any consistent offset that might be due to a single systematic factor such as miscalibration of the clinic display monitor.

Array of Digital Images by B/C/CB Parameters
To illustrate the range of B/C/CB parameters in AREDS2 digital images, Figure 5 presents an array of digital images organized by level of G/R color balance (rows) and by percentile of brightness (columns). Each cell of this cross-tabulation is populated by a representative image selected from the candidate pool with those parameters, favoring those containing obvious abnormalities of nonadvanced AMD and appearing near the middle of that category’s range. Luminance histograms are included so that the reader can correlate the appearances of the images and their histograms.


Figure 5
View larger version (65K):
[in this window]
[in a new window]

  		FIGURE 5. Array of representative AREDS2 digital color fundus images (sample n = 605 eyes), based on the better side of the right eye macular stereo pair. Rows present a progression of G/R color balance ratio (calculated by comparing the locations of the two curve peaks in the dynamic range) from 0.35 to 0.50, and columns display a progression of brightness (determined by the location of the red curve peak in the dynamic range) from 10th through the 95th percentiles. For each brightness percentile, the red curve peak location is specified in 16ths of the dynamic range. For comparison, a series of representative film images (n = 196 eyes) is included as the top row, restricted to those with G/R ratio of 0.50, progressing through percentiles of brightness. Note variation in visibility of typical AMD abnormalities (drusen and RPE pigment disturbances) in images with different B/C/CB.

 
Rows are included for G/R color balance (G/R ratio) of 0.50 for both film images (for which this value was both median and mode) and digital (for which this was the mode). In addition, digital has a row for G/R ratio of 0.35 (most discrepant from model image color balance). Among digital images, 23% were in this markedly reddish category, compared to 9% for film (data not shown). Regarding brightness, images at the 10th percentile were substantially darker for digital than film (red peak at 6/16 vs. 10/16). Digital images did not match the target brightness from the model until the 75th percentile (12/16), whereas film images did by the 50th percentile (13/16). Both digital and film images at the 95th percentile (red peak at 16/16) had oversaturation of the red channel. Such images are easily identifiable from their histograms, as illustrated by the far right histograms: the red curve is truncated or "clipped" at the top (right) end of the dynamic range, showing a spike at the white point (representing the quantity of pixels saturated) rather than a normal right tail. Defining red channel oversaturation pragmatically as 15% or more of the pixels in the top 16th of the dynamic range, 19% of digital images versus 4% of film images were overilluminated (data not shown).

Figure 5 illustrates various combinations of B/C/CB parameters identified in Table 2 and Figure 4 . The second column shows images with median overall brightness (keying on the red channel) and G/R color balance ratio (0.50)—top row for film and bottom row for digital. Note that the median film image tends to be somewhat brighter than the median digital image. A comparison of film and digital images in the first column, which shows the dimmest images, showed that film tends to be brighter than digital. However, a comparison of film and digital images in the last column, which shows the brightest (overexposed) images, revealed that the detrimental effect was similar in both modes.

Summary of Tonal Resolution Differences between Digital and Film
Table 3 summarizes the comparison between film and digital images, keying on the crucial differences in the range and distribution of the B/C/CB parameters. Problematic characteristics were two- to many-fold more prevalent in digital images.


View this table:
[in this window]
[in a new window]

  		TABLE 3. AREDS2 Digital versus Film Color Fundus Images: Frequency of Unsatisfactory B/C/CB

 
Simulation of Tonal Variation in a Single-Color Fundus Image
Although the array of representative digital images in Figure 5 allows the observer to form an overall sense of the varying suitability of different tonal parameters for documentation of nonadvanced AMD, exact comparisons are not feasible because these photographs are all different eyes from different subjects. Using image analysis software (Photoshop; Adobe Systems), we manipulated the tonal parameters of one image with AMD content (from Fig. 3 ) to produce Figure 6 , an array of images organized by gradations of brightness and color balance as in Figure 5 . Color balances between 0.50 (the model value) and 0.35 (weak green) were inserted—0.45 as closest to the median of AREDS2 digital images and 0.40 to show the effect of weaker green that is less extreme.


Figure 6
View larger version (77K):
[in this window]
[in a new window]

  		FIGURE 6. Array simulating the effect on a single image with AMD content of the progression shown in Figure 5 of the G/R color balance ratio (calculated by comparing locations of the two curve peaks in the dynamic range) and percentile of brightness (determined by the location of the red curve peak in the dynamic range). Histograms at the bottom of each column are a composite representing both the top and bottom images in that row. Red and blue parameters are the same: The shaded green curve lower in the range is the position for G/R ratio of 0.35, and the solid green curve higher in the range is the position of G/R ratio = 0.50. Note the difference in the contrast of AMD abnormalities against RPE background as brightness and color balance is shifted.

 
Impact of the Blue Component in Color Fundus Images
Figure 7 shows the impact of different levels of blue in the color fundus image, using the approaches of Figure 5 (a representative digital series from AREDS2, and a film series for comparison) and Figure 6 (a simulated series using one image). As the B/R ratio exceeded the model value of 0.17, the image took on a purplish cast that Reading Center graders perceived as decreasing contrast of AMD abnormalities against the normal RPE background. (This is a subjective effect—a higher blue level cannot alter the objectively measured contrast in the red and blue channels.) Note that the absolute B/R ratio at the 75th percentile for film remained below the 50th percentile value for digital, indicating stronger influence from the relatively uninformative blue channel in some digital images.


Figure 7
View larger version (75K):
[in this window]
[in a new window]

  		FIGURE 7. Series displaying progression of B/R color balance ratio (calculated by comparing locations of the two curve peaks in the dynamic range), shown both with representative digital images (n = 605 eyes) as in Figure 5 and with simulations constructed from the same image as in Figure 6 . Image brightness (determined by the location of the red curve peak in the dynamic range) is fixed at the 75th percentile and G/R color balance at 0.50. For comparison, a series of representative film images (n = 196 eyes) constitutes the top row. Note any decrease in apparent contrast of AMD abnormalities against RPE abnormalities as B/R color balance is shifted higher.

 
Applying the Model Image as a Basis for Standardized Enhancement
After analyzing AREDS2 images with luminance histograms, graders used digital enhancement, guided by algorithms, to adjust the B/C/CB parameters of substandard images to match those of the iMD Chrome histogram model. Figure 8 shows the standardized enhancement of several types of problematic images, including the originals for comparison. Optimization was performed by applying adjustment tools to achieve the target spans and locations separately for red, green, and blue luminance curves. Each color channel from each individual image may require contrast expansion/compression and/or brightness increase/decrease. (Production enhancement of AREDS images used IMAGEnet, restricted to sliders that adjust black and/or white points. The more extreme images in Figure 8 were adjusted with Photoshop, which allows precise control via keyboard entry of variables.) In short, the curves were modified to their target span and moved to their target location, as dictated by the image model. (Optimization of the blue channel is impossible if that channel is extremely compressed or absent in the original image.) Examination of the before and after color histograms demonstrates the standardization of B/C/CB parameters.


Figure 8
View larger version (76K):
[in this window]
[in a new window]

  		FIGURE 8. Standardized optimization of B/C/CB of obviously problematic digital color fundus images from the AREDS2 sample, displaying before and after versions with histograms. Image enhancement was performed according to the "recipe" described as iMD Chrome in Table 2 . General cases are those in which the image can be fully enhanced to match the parameters of the image model: (A) too dark, but proper color balance; (B) excessively red; (C) excessively green; and (D) excessively blue. Special cases are those in which the image cannot be fully enhanced due to loss of pictorial information in the original: (E) over-saturated in the red channel. The red channel breakout is shown in monochrome to highlight resultant loss in RPE detail.

 
Figure 8A presents the general case: an underilluminated image, but with reasonable color balance. Figures 8B 8C 8D present variants of the general case: excessively red (but not clipped; Fig. 8B ), green (Fig. 8C) , and blue (Fig. 8D) images. Figure 8E shows an image with an oversaturated (clipped) red channel—a special case because image information had been irretrievably lost so that only partial restoration was possible. The best result was obtained by shifting the remnant of the red curve down to its standard position and adjusting the other two curves in standard fashion.

Figure 9 presents the results of standardized optimization on the image row considered most problematic in Figure 5 —that with G/R color balance ratio of 0.35 (red too strong and green too weak). Percentiles of brightness from the 10th through the 95th have been included to show the standardizing effect of optimization.


Figure 9
View larger version (50K):
[in this window]
[in a new window]

  		FIGURE 9. Standardized enhancement of B/C/CB applied to the top row of digital images from Figure 5 (having variable illumination and improper G/R color balance). Before and after versions are shown in the top and bottom rows, with histograms. Enhancement algorithm implemented the image "recipe" described as iMD Chrome in Table 2 . Compare the contrast of typical AMD abnormalities against the RPE background in the original versus optimized images.

 
AREDS2 Reading Center graders rated images for photograph quality after optimization, taking into consideration not only the historic criteria of field definition, focus/clarity, and stereoscopic effect, but also brightness, contrast, and color balance. Comparing AREDS2 quality statistics from the first internal study report (n = 653 eyes, 3:1 mixture of film and digital images) with those from the original AREDS (n = 11,174 eyes, all film images) yielded the following results: satisfactory images, 95.6% vs. 94.4%; acceptable/borderline, 4.4% vs. 5.0%; and ungradable, 0.0% vs. 0.6%, respectively. With optimization of B/C/CB parameters, AREDS2 has been able to sustain a level of image quality that graders judged to be consistent with that of its predecessor.


   Discussion
Top
 Abstract
 Population and Methods
 Results
 Discussion
Conclusions
 References
 
Before AREDS2 began, at least three other groups using similar imaging and grading procedures had reported satisfactory comparability between AMD gradings of digital and film images using related classifications.3 4 5 Therefore, we did not consider it necessary to repeat that exercise. However, all these successful studies were performed taking both digital and film images within the same subjects in single clinics. Although Klein et al.4 and Somani et al.5 did not describe how they adjusted their digital cameras, van Leeuwen et al.3 conducted a pilot study during which they calibrated their digital camera for B/C/CB based on side-by-side comparison with their film images.

Given the capabilities and preferences of the 90 clinical centers participating in AREDS2, for practicality we had to allow most clinics to switch to digital. In a multicenter clinical trial, it is imperative to standardize documentation of the fundus as much as possible. Furthermore, because we intend to perform meta-analyses across AREDS2 and film-based studies—AREDS and large epidemiologic studies such as the Beaver Dam Eye Study8 —it is important to understand how methodological differences may impact results and to minimize differences as far as possible.

By constructing an image model based on the shared characteristics of excellent quality images selected by different expert groups, we took advantage of an observation by novelist Robert Pirsig9 : people may not agree on a definition of quality, but they often agree on examples of it. Histograms of image exemplars reveal remarkable consistency in B/C/CB parameters, because the retinal scene is very predictable, with normal content accounting for most of the pixels. In other words, disease abnormalities typically have little impact on the robust color curves generated by most fundus images. An image with exemplary B/C/CB generates a histogram somewhat resembling a suspension bridge: the red and green peaks, located in the upper and lower halves of the dynamic range, represent the towers, and their curves represent the cables, overlapping slightly in the middle (blue makes a minor appearance near the bottom). (Readers with access to Photoshop can try the proposed enhancement procedure by loading their own color retinal images into that program, invoking the Image/Adjustment/Levels tool, and then using the black and white point sliders, separately for red, green and blue channels, to manipulate the curves to approximately their target locations and spans.)

Suitability of the iMD Chrome model for images taken to document AMD was demonstrated pragmatically. We found that measured contrast between typical AMD abnormalities and normal RPE background was maximized when specimen images were tuned to the model.

A comparison of digital to film fundus images in AREDS2 showed that the quality of the best digital images matched that of the best film, in our opinion. However, histogram analysis (Fig. 4) and visual inspection (Fig. 5) of large samples revealed that digital had wider variability in its tonal resolution than did film. There were more outliers at both ends of the spectrum—dark images with low contrast, and oversaturated images that, paradoxically, also displayed low contrast.

AREDS2 Reading Center graders subjectively prefer images that are moderately bright (50th–75th percentile) with higher G/R ratio (0.50–0.55), which are perceived as having optimum contrast between AMD abnormalities, such as drusen and pigment abnormalities, against the RPE background. On luminance histograms, such images manifest broad red and green curves. Images at the margins of the array are considered less satisfactory because those at or below the 10th percentile for brightness appear too dim, those at or above the 90th percentile too bright, and those with color balance below 0.45 too red-saturated.

There are fundamental reasons why the digital medium, given current practice, inherently has more variable B/C/CB management. (1) The retina presents a narrow color gamut compared with most other scenes, with red reflectance being brightest, even though most spatial detail occurs in the green. (2) Digital cameras are constructed with silicon CCD sensors especially responsive to red, resulting in images that tend to have strong red and weak green. (3) Compared with film emulsions, digital sensors have a narrow dynamic range, making it more difficult for the photographer to find the best illumination level. (4) The response profile of the digital sensor is linear, whereas that of film is curved. In film, the red response is damped as illumination approaches the maximum, whereas in digital it quickly produces oversaturation. (The digital combination of narrow dynamic range and linear response often results in oversaturation of the disc, masking abnormalities such as new vessels. Possible solutions, such as variable image enhancement or remapping of digital sensor response, are beyond the scope of this report, which focuses on AMD.) (5) Manufacturers do not always provide clear recommendations for the proper camera settings, and even when they do, the systems may not remain in adjustment over time. Settings that regulate B/C/CB may not be readily accessible to the photographer, and even when they are, the photographer may not know how to adjust them. (6) The photographer may be judging images from their appearance on an uncalibrated monitor, or even from a hard-copy print.

Putting these factors together, we suspect that the most critical factor is color balance. If the camera is set for strong red and weak green (e.g., G/R ratio below 0.45), the photographer increases the flash looking for fundus detail to emerge clearly, but does not see it so keeps going until the red channel oversaturates. (Conversely, photographers encountering red oversaturation may keep the flash dim thereafter to avoid it.) If the camera is set for balanced G/R, the photographer sees the desired fundus detail emerge as the flash is increased, and thus is able to capture intuitively a well-illuminated picture that avoids extremes of under- and overillumination.

Reading center interaction with photographers struggling with B/C/CB problems suggests that, in order for them to obtain the best results, the following are necessary: (1) the ophthalmic community must develop some consensus on criteria for B/C/CB parameters; (2) digital camera manufacturers should specify their recommended settings, so that photographers have a sound starting position from which to make any further adjustments necessary for their particular systems; (3) monitors (and printers) must be profiled and calibrated for accurate display; (4) photographers, ophthalmologists, and researchers must have sufficient "digital image literacy" to be conversant with B/C/CB issues; (5) digital cameras may need to allow as many as three saved color balance settings—one for light pigmentation, one for medium, and one for dark; and (6) digital cameras must have tools in the capture interface to allow photographers to analyze and manage tonal parameters for individual patients.

However, at the current stage of digital imaging technology, we suspect that most photographers will not be able to obtain the most optimal images at the time of capture, even with best practice. Therefore, we think that post hoc image enhancement is necessary to obtain the best results in many instances, whether done by the photographer, the ophthalmologist, or the reading center. As shown in Figures 8 and 9 , the optimized images display a much more uniform appearance than do the original digital images, and attain a consistency of tonal resolution that makes them more comparable to film.

Possible weaknesses of this approach are the following: (1) Our iMD Chrome fundus image model was constructed specifically for AMD. Although we think it has some advantages for other major retinal diseases our Reading Center studies (e.g., for diabetic retinopathy the strong green content sharpens contrast of blood vessels, microaneurysms, IRMA (intraretinal microvascular abnormalities), and new vessels against the RPE), this image recipe may not be suitable for all other purposes. In particular, our method needs further development to avoid frequent oversaturation of the disc, which could mask new vessels. (2) Our optimization is oriented toward subjective appreciation of images—objective image processing (e.g., automated drusen detection) may have more exacting requirements. (3) Standardized optimization suppresses individual differences in level of RPE pigmentation, in favor of enhancing the contrast of AMD abnormalities against the RPE background. (Because neo-AMD is less frequent in African-Americans, they are underrepresented in AREDS2, limiting our ability to address this topic.) For study of differences in pigmentation level, this standardization would be undesirable. (4) Our characterization of curve location and span (brightness and contrast) using tails and peaks may not be as accurate as a more sophisticated mathematical analysis of these quasi-Gaussian curves. (5) The RGB color space may not be best for rebalancing the B/C/CB parameters of digital images, compared to other spaces such as HSV (hue, saturation, value) LAB (L for luminance, A and B the color opponent dimensions), or YUV (one luma component, two chrominance components). Concerning the last two points, our simple approach had to be practicable for manual adjustment—an automated processor might be able to take advantage of more powerful tools.

By enhancing images in AREDS2, we are modifying source documents—albeit in a standardized manner supported by an explicit rationale. As clinical trials transition to digital images, standardization of parameters by controlling film emulsion and development has been unintentionally abrogated. To illustrate, Figure 10 displays images of the same normal eye taken on six different makes of digital camera, each used in the manner customary for the clinic in which it was located. All of these images are source documents, but which one of the various appearances is true? Ultimately, the eye itself is the actual source, and the original photograph taken of it is already impacted by a host of highly variable factors. Thus, measures to reimpose standardized B/C/CB parameters could be considered due diligence for a multicenter clinical trial or long-term epidemiologic study depending on digital color fundus images for its outcomes.


Figure 10
View larger version (42K):
[in this window]
[in a new window]

  		FIGURE 10. Color fundus images of the same normal left eye from six different masked makes of fundus camera (A–F), each adjusted and used as customary for that clinic (not necessarily as recommended by the manufacturer). (Courtesy of Christye Sisson, MS, CRA, Rochester Institute of Technology, Rochester, NY.) Note the variety of B/C/CB parameters across the six different images. Optimized versions according to the iMD Chrome model are displayed beneath the originals.

 

   Conclusions
Top
 Abstract
 Population and Methods
 Results
 Discussion
 Conclusions
References
 
The AREDS2 approach to analysis and management of B/C/CB in fundus images appears to optimize their utility for documentation of AMD abnormalities. We developed B/C/CB criteria (expressed by a model image histogram) and used algorithms to enhance images accordingly. In the process, we correlated the subjective appearance of images and their objective histograms to improve our image literacy. Our effort is only part of what must necessarily be a larger movement to improve the quality of digital images, relying on the ophthalmic photographers and their professional organization, the digital camera manufacturers, and ultimately the ophthalmologists who use these images in daily practice for patient diagnosis and care.


   Footnotes
 
Supported by Contract HHSN 260-2006-00003C, National Eye Institute, Bethesda, MD.

Submitted for publication September 27, 2007; revised November 26, 2007, and January 11, 2008; accepted June 18, 2008.

Disclosure: L.D. Hubbard, None; R.P. Danis, None; M.W. Neider, None; D.W. Thayer, None; H.D. Wabers, None; J.K. White, None; A.J. Pugliese, Choices for Service in Imaging, Inc. (E, P); M.F. Pugliese, Choices for Service in Imaging, Inc. (E, 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.

Corresponding author: Larry D. Hubbard, University of Wisconsin Fundus Photograph Reading Center, 406 Science Drive, Suite 400, Madison, WI 53711-1068; hubbard{at}rc.ophth.wisc.edu.


   References
Top
 Abstract
 Population and Methods
 Results
 Discussion
 Conclusions
 References
 

  1. Age-Related Eye Disease Study Group. The Age-Related Eye Disease Study system for classifying age-related macular degeneration from stereoscopic color fundus photographs. Am J Ophthalmol. 2001;132:668–681.[CrossRef][Web of Science][Medline][Order article via Infotrieve]
  2. Age-Related Eye Disease Study Group. The Age-Related Eye Disease Study severity scale for age-related macular degeneration. Arch Ophthalmol. 2005;123:1484–1498.[Abstract/Free Full Text]
  3. Van Leeuwen R, Chakrvarthy U, Vingerling JR, et al. Grading of age-related maculopathy for epidemiological studies: is digital imaging as good as 35-mm film?. Ophthalmology. 2003;110:1540–1544.[CrossRef][Web of Science][Medline][Order article via Infotrieve]
  4. Klein R, Meuer SM, Moss SE, et al. Detection of age-related macular degeneration using a nonmydriatic digital camera and a standard film fundus camera. Arch Ophthalmol. 2004;122:1642–1646.[Abstract/Free Full Text]
  5. Somani R, Tennant M, Rudnisky C, et al. Comparison of stereoscopic digital imaging and slide film photography in the identification of macular degeneration. Can J Ophthalmol. 2005;40:293–302.[Web of Science][Medline][Order article via Infotrieve]
  6. Scholl HPN, Dandekar SS, Peto T, et al. What is lost by digitizing stereoscopic fundus color slides for macular grading in age-related maculopathy and degeneration?. Ophthalmology. 2004;111:125–132.[CrossRef][Web of Science][Medline][Order article via Infotrieve]
  7. Lee MS, Shin DS, Berger JW. Grading, image analysis, and stereopsis of digitally compressed fundus images. Retina. 2000;20:275–281.[CrossRef][Web of Science][Medline][Order article via Infotrieve]
  8. Klein RK, Klein BEK, Tomany SC, Meuer SM, et al. Ten-year incidence and progression of age-related maculopathy. Ophthalmology. 2002;109:1767–1779.[CrossRef][Web of Science][Medline][Order article via Infotrieve]
  9. Pirsig RM. Zen and the Art of Motorcycle Maintenance. 1974; William Morrow and Co. New York.



This article has been cited by other articles:


Home page
 Arch OphthalmolHome page
The AREDS Research Group
Change in Area of Geographic Atrophy in the Age-Related Eye Disease Study: AREDS Report Number 26
Arch Ophthalmol, September 1, 2009; 127(9): 1168 - 1174.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
iovs.07-1267v1
49/8/3269    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Web of Science (2)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Hubbard, L. D.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Hubbard, L. D.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS