Abstract
Purpose.:
Changes in retinal vascular caliber measured from digital color fundus photographs have been independently associated with systemic outcomes in epidemiologic studies, but the effect of image resolution and compression on vascular measurements has not been previously evaluated.
Methods.:
To explore image compression, 40 natively digital fundus images were selected with good photo quality, high spatial resolution, and no previous image compression. Using Adobe Photoshop, these images were compressed at progressively higher levels up to 147:1, and then retinal vascular caliber was measured at each level using semiautomated software. To examine resolution, 40 fundus photographs acquired on high-resolution film were scanned with settings corresponding to 10, 7, 5, 3, and 1 megapixel fundus cameras. After adjusting for scale factor, vascular caliber was measured at each level of resolution. Data were analyzed by comparing the calculated central retinal arteriole equivalent (CRAE) and the central retinal venular equivalent (CRVE) of the original and altered images, using repeated measures ANOVA.
Results.:
CRAE became significantly wider with increasing levels of compression at the 25:1 threshold (∼1 μm wider, P < 0.001) and was ∼5 μm wider with 147:1 compression. CRVE also increased, but less than CRAE. Using 7 (megapixel)-MP resolution as the standard, CRVE was significantly narrower at the 5-MP simulation (∼2 μm, P < 0.001) and was ∼12 μm narrower at the 1-MP simulation. CRAE also decreased, but less than CRVE.
Conclusions.:
Increasing digital image file compression and decreasing fundus image spatial resolution led to skewed measurements of the retinal vascular caliber.
Epidemiologic studies have reported that retinal vascular caliber changes are independent risk factors for different systemic health outcomes. Collectively, this methodology has been applied to the images of more than 40,000 individuals.
1 Commonly, retinal vascular caliber is measured from digital color fundus images—natively digital or scanned from film—using specialized software that detects vessel edges in order to calculate the width of arterioles and venules. For each eye, these individual vessel widths are then combined using a generalized model (Knudtson revision of the Parr-Hubbard formulas) to obtain the central retinal arteriolar equivalent (CRAE) and the central retinal venular equivalent (CRVE).
2
The uncompressed files of digital images may be large—on the order of 14 MB for a 5- MP camera. To facilitate storage and transportation of large images, they are often compressed to reduce file size. However, any compression beyond ∼4:1 requires the use of “lossy” (as opposed to lossless) compression, typically in the joint picture expert group (JPEG) format. Such compression increasingly tends to diminish image detail as the file size decreases. Thus, it is critically important to know how much the image is compressed as the file size decreases, to ensure that sufficient image integrity is preserved for accurate measurement of the retinal vessels.
Closely related is the issue of spatial resolution—in digital images the x × y pixel dimensions of the image, expressed in megapixels. The color slide transparency emulsion used historically, Ektachrome or equivalent (Eastman Kodak Company, Rochester, NY), has been estimated to correspond to the spatial resolution of a 10-MP digital image, approximately 3000 × 3600 pixels.
3 Regardless of digital image compression, various digital fundus camera systems capture images at different pixel dimensions, which also impacts file size. Two primary parameters of digital fundus cameras that affect the spatial resolution of the digital image are the degree of magnification of the image provided by the camera optics, and the dimensions of the pixel array of the digital sensor itself. However, the size of the sensor pixel array available to capture the image is fixed regardless of the optical field setting. Thus, from any given camera, wide-angle images at 45° to 50° have lower spatial resolution than do approximately 30° images. Because the variety of digital fundus cameras and digital sensors now in use is greater than the variety of historical film cameras, it is prudent to be cognizant of the possible impact of spatial resolution of the digital fundus image on retinal vascular caliber measurements.
To our knowledge, progressive image compression has not been studied in the context of retinal vascular caliber measurement. Nor are we aware of any research on the possible impact of differential image resolution on accuracy of retinal vascular measurement.
Images from 40 participant eyes were randomly selected from multiple studies within the University of Wisconsin-Madison Fundus Photograph Reading Center database. These studies had been conducted in accordance with the tenets of the Declaration of Helsinki and under supervision of the respective institutional review boards. All participants had given written informed consent to have their fundus photographs taken for research purposes. Using Irfanview image handling software (Version 4.27; Irfan Skiljan, Vienna, Austria), the image information was viewed to ensure that the image had no previous compression. Furthermore, the image evaluator had to deem the image focus and illumination to be adequate. All photos had been taken with digital cameras manufactured by either Topcon (Topcon Medical Systems Inc., Oakland, NJ) or Zeiss (Carl Zeiss Meditec Inc., Dublin, CA). The average file size of the original uncompressed tagged image file format (TIFF) image was approximately 14 MB.
For vascular measurement, the “scaling” factor or the “pixel spacing” parameter (the Digital Imaging and Communications in Medicine [DICOM] standard; National Electrical Manufacturers Association, Rosslyn, VA, 2009), of each image is determined prior to evaluation. The number of pixels across a known distance (disc to macula) is evaluated, and the pixel spacing parameter is calculated by dividing the number of pixels by the known distance. In this study, while the original image is down-sampled for analysis, the pixel spacing parameter is unchanged by image compression but altered as spatial resolution changes. We used the landmarks of the center of the disc and the center of the macula, which is presumed to be 4500 μm, to determine the pixel spacing value.
Predominantly, retinal imaging in clinical settings uses JPEG compression.
4 Accordingly, for this exercise the uncompressed TIFF images were imported into Adobe Photoshop (CS3 V. 10.0; Adobe Systems Inc., San Jose, CA) and then converted to JPEG images at compression ratios ranging from 4:1 to 147:1. The compressed JPEG files were then converted back to TIFF format as required to use the vessel measurement software, IVAN (Version 1.3; Nicola Ferrier, Madison, WI). For each camera and angle combination, an average pixel spacing factor was determined across all the original images. Image compression does not change the pixel spacing factor, thus the same factor is applicable at all compression ratios. To eliminate interobserver variability, one person evaluated all the images in this study, with the original image being evaluated first. Before recording the data of the original image, a screenshot was saved as a reference for grid placement and vessel identification, ensuring that the corresponding sections of the vessels would be measured on the subsequent compressed images. After summarization into the CRAE and CRVE, results were collected for the statistical analysis, which is described later. A flow chart of the methods used for the compression portion of the study is shown in
Figure 1.
The effect of progressively lower spatial resolution was assessed by analyzing measurements on images that correspond with 10-, 7-, 5-, 3-, and 1-MP cameras. To simulate different camera resolutions using the same images, we used film images from 40 eyes of participants who were randomly selected from the multiple studies within the University of Wisconsin-Madison Fundus Photograph Reading Center database. The camera types from the images selected were the Zeiss 30, Topcon 50, and Canon 60 degree film cameras (Canon U.S.A., Inc., Lake Success, NY).
Slide images were digitized using a Nikon Co-olscan V ED 1.02 scanner (Nikon Inc., Melville, NY) at the pixel dimension calculated to represent the different camera spatial resolutions listed previously. Changing the resolution of an image alters the pixel spacing parameter; therefore, we had to establish the corresponding pixel spacing parameter for each subsequent resolution. As each series was created by altering an original slide, the millimeter distance from the center of the disc to the center of the macula remained unchanged, as the only factor varying through the series was the pixel spacing parameter in each image representing this distance. With this knowledge, a calibration ratio was created that could be used to scale each image relative to others in the set. Therefore, by measuring the distance between the center of the disc to the center of the macula and the distance across the entire image in pixels, the ratio can be calculated for each image. The average ratio from the five images was used to calculate the calibration ratio for that image set. The calibration ratio was then implemented by multiplying it against the measured distance across each image. This yielded a disc to macula measurement that was consistent across all spatial-resolution images in millimeters regardless of pixel spacing, ensuring accurate measurement and grid fit. The 7-MP images were evaluated first, on a monitor with a display resolution of 1600 × 1200 pixels, with the pixel spacing changed accordingly within the configuration file before evaluating the other images. Similar to the compression portion of this exercise, we saved a screenshot of the original image to use as a reference for subsequent images. A flow chart of the methods for the spatial resolution portion of the study is shown in
Figure 2.
In both studies, CRAE and CRVE results were compared by calculating the signed difference between the original images and each image variant in the study. In analysis of variance for repeated measures design, each image serves as its own control, and accordingly, variability owing to differences in average responsiveness of the images is eliminated from the extraneous error variances. Data management and analyses were performed in SAS (SAS/STAT software, Version 9.2; SAS Institute Inc., Cary, NC).
Sample size calculations were performed using G*Power (Version 3.1.3; Franz Faul, Kiel University, Kiel, Germany). The sample size required to demonstrate a difference within two compression levels was 27 images (α = 0.05, power = 0.80, and effect size = 0.1), while the sample size for resolution was 38 images (α = 0.05, power = 0.80, and effect size = 0.1).
Retinal vascular caliber is routinely measured at the reading center for other observational cohorts, and we perform reproducibility exercises as part of the quality control process. Historically we have observed that variability within evaluator, the relevant consideration in this experiment, is less than that between evaluators.