Article Title:

Screening using compressed digital retinal images successfully identifies retinopathy.

Back to Published Articles



Diabetes Care. 2003 Jan;26(1):247.


Graham P. Leese, MD, FRCP, Angela Ellingford, BSC, Andrew D. Morris, MD, FRCP, John D. Ellis, MPH, FRCOphthal and Scott Cunningham, BSC


Dr. Graham Leese, Ward 1 and 2 Ninewells Hospital, Dundee DD1 9SY, U.K. E-mail:


Digital retinal photographs can be integrated into a computerized network that more easily enables communication and quality assurance (1). Image compression overcomes the difficulty of transmitting and storing large file sizes. Concern has been raised that major compression of 70% results in clinically significant loss of retinal detail with inadequate screening sensitivities (2). It is unclear whether low levels of compression result in loss of screening sensitivity compared with the original bit-map image.

We used a Topcon TRC-NW6S nonmydriatic fundus camera with a Sony DXC950P to photograph 171 patients with diabetes (one eye each for the study), without the use of mydriasis. Original bit-map images (768 x 576 pixels, 1.27 MB) were stored and compressed to make a JPEG image (104 KB) of the highest quality using Paintshop Pro (Jasc Software, Eden Prairie, MN) with standard encoding. All images were anonymized and presented to the grader in random order. Images were graded on a 17-inch Cathode ray tube monitor with 1,024 x 768 pixel resolution in a darkened room by a single grader. Severe and very severe nonproliferative retinopathy, proliferative retinopathy, and maculopathy were defined as vision-threatening retinopathy.

On the original bit-map images, 80 patients had normal retina (46.7%), 35 had background retinopathy (20.5%), 38 had vision-threatening retinopathy (22.2%, 5 proliferative and 33 maculopathy), 8 had non-diabetes-related changes (4.7%), and 10 were unreadable (5.8%). Compared with bit-map images, grading using the JPEG images achieved a sensitivity of 95.8% (±5.1%, 95% CI) and a specificity of 95.0% (±4.2%) in the detection of any identifiable disease. This yields a positive predictive value of 94.6% and a negative predictive value of 96.2%. In terms of identifying vision-threatening retinopathy, the sensitivity of using highest-quality compressed JPEG images was 97.4% (±2.4%) with a specificity of 100%. The positive predictive value was 100%, and the negative predictive value was 99.3%. The difference between JPEG images and bit-map images in the detection of vision-threatening referable disease amounted to a disagreement about the presence of one microaneurysm in one image, which did not require subsequent laser photocoagulation.

Using highest-quality compressed JPEG images (Paintshop Pro) does not appear to result in any loss of sensitivity when compared with uncompressed bit-map images for detecting potentially vision-threatening disease. This finding helps confirm earlier pilot studies (3,4). JPEG files compressed to highest-quality images result in file sizes that are 8% of the original bit-map image file size, which allows them to be more readily stored, more easily transferred across a web-interface, and transmitted at a faster rate.