Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2009 Jul 1;3(4):819-34.
doi: 10.1177/193229680900300431.

Screening for diabetic retinopathy using computer vision and physiological markers

Christopher E Hann  1 , James A Revie, Darren Hewett, J Geoffrey Chase, Geoffrey M Shaw
Affiliations

Screening for diabetic retinopathy using computer vision and physiological markers

Christopher E Hann et al. J Diabetes Sci Technol. .

Abstract

Background: Hyperglycemia and diabetes result in vascular complications, most notably diabetic retinopathy (DR). The prevalence of DR is growing and is a leading cause of blindness and/or visual impairment in developed countries. Current methods of detecting, screening, and monitoring DR are based on subjective human evaluation, which is also slow and time-consuming. As a result, initiation and progress monitoring of DR is clinically hard.

Methods: Computer vision methods are developed to isolate and detect two of the most common DR dysfunctions-dot hemorrhages (DH) and exudates. The algorithms use specific color channels and segmentation methods to separate these DR manifestations from physiological features in digital fundus images. The algorithms are tested on the first 100 images from a published database. The diagnostic outcome and the resulting positive and negative prediction values (PPV and NPV) are reported. The first 50 images are marked with specialist determined ground truth for each individual exudate and/or DH, which are also compared to algorithm identification.

Results: Exudate identification had 96.7% sensitivity and 94.9% specificity for diagnosis (PPV = 97%, NPV = 95%). Dot hemorrhage identification had 98.7% sensitivity and 100% specificity (PPV = 100%, NPV = 96%). Greater than 95% of ground truth identified exudates, and DHs were found by the algorithm in the marked first 50 images, with less than 0.5% false positives.

Conclusions: A direct computer vision approach enabled high-quality identification of exudates and DHs in an independent data set of fundus images. The methods are readily generalizable to other clinical manifestations of DR. The results justify a blinded clinical trial of the system to prove its capability to detect, diagnose, and, over the long term, monitor the state of DR in individuals with diabetes.

PubMed Disclaimer

Figures

Figure 1.
Figure 1.
Exudates and optic disk.
Figure 2.
Figure 2.
(A) Result of choosing the top 0.5% of the green intensities. (B) Locating the optic disk based on the largest connected region. The region is surrounded by a circle.
Figure 3.
Figure 3.
Image 029 from the database with a blue line through x = 153 and an exudate pointed out with an arrow.
Figure 4.
Figure 4.
(A) Plotting the green intensity along the line shown in Figure 3. (B) Result of subtracting the 50-pixel median from each intensity.
Figure 5.
Figure 5.
(A) Example of applying median filter method of Figure 4 and (B) result after adding pixels.
Figure 6.
Figure 6.
(A) Selecting contours of the red/green surface for a true exudate. Candidate contours are dashed lines, exudate is denoted by circles, and blue solid line is the final chosen contour. (B) Contours of red/green surface for false exudate.
Figure 7.
Figure 7.
Algorithm for detecting exudates.
Figure 8.
Figure 8.
Fundus image with dot hemorrhage and fovea identified.
Figure 9.
Figure 9.
Image of red/green ratio intensities.
Figure 10.
Figure 10.
Scaling using a median filter on the red/green ratio in the vertical direction for one column of the image in Figure 9. No units are given as the y axis is a ratio, which cancels out the units.
Figure 11.
Figure 11.
Binary image after thresholding.
Figure 12.
Figure 12.
Illustration of calculating the shape number for a connected region.
Figure 13.
Figure 13.
Simplified calculation of the shape number for a connected region.
Figure 14.
Figure 14.
Calculating the shape number for (A) a vessel, (B) a dot hemorrhage, and (C) the fovea.
Figure 15.
Figure 15.
Selection of round regions.
Figure 16.
Figure 16.
Results of minimum distance lines found from Delaunay triangulation.
Figure 17.
Figure 17.
Red/green ratio intensity along lines 1 and 2.
Figure 18.
Figure 18.
Dot hemorrhage segmentation from the fundus image.
Figure 19.
Figure 19.
Summary of dot hemorrhage detection algorithm.
Figure 20.
Figure 20.
(A) Image 002, a darker image. (B) Results of the algorithm of Figure 7.
Figure 21.
Figure 21.
(A) Image 005. (B) Results of the algorithm of Figure 7 on image 005 showing that the actual number of exudates is captured accurately.
Figure 22.
Figure 22.
(A) An example of applying the algorithm of Figure 7 to image 005 in Figure 21 without the contour checker. (B) Figure 7 applied to image 099 without the contour checker.
Figure 23.
Figure 23.
(A) A false positive (pointed to by an arrow) found by the algorithm of Figure 7 in image 64. However, a potential exudate can be seen. (B) A false negative corresponding to image 71. Dot hemorrhages can be seen, but it is hard to determine the existence of exudates visually.
Figure 24.
Figure 24.
Fundus image of false negative.
See this image and copyright information in PMC

Similar articles

See all similar articles

References

    1. Icks A, Trautner C, Haastert B, Berger M, Giani G. Blindness due to diabetes: population based age- and sex- specific incidence rates. Diabet Med. 1997;14(7):571–575. - PubMed
    1. Fong D, Aiello L, Gardner T, King G, Blankenship G, Cavallerano J, Ferris F, 3rd, Klein R. Retinopathy in diabetes. Diabetes Care. 2004;27(Suppl 1):S84–S87. - PubMed
    1. Kempen JH, O'Colmain BJ, Leske MC, Haffner SM, Klein R, Moss SE, Taylor HR, Hamman RF. Eye Diseases Prevalence Research Group. The prevalence of diabetic retinopathy among adults in the United States. Arch Ophthalmol. 2004;122(4):552–563. - PubMed
    1. Wild S, Roglic G, Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the year 2000 and projections for 2030. Diabetes Care. 2004;27(5):1047–1053. - PubMed
    1. Fong D, Aiello L, Gardner T, King G, Blankenship G, Cavallerano J, Ferris FL, 3rd, Klein R. American Diabetes Association. Diabetic retinopathy. Diabetes Care. 2003;26(1):226–229. - PubMed