May 2006
Volume 47, Issue 13
ARVO Annual Meeting Abstract  |   May 2006
A Statistical Approach to Automatic Screening for Diabetic Retinopathy
Author Affiliations & Notes
  • G. Zamora
    Biomedical, Kestrel, Albuquerque, NM
  • S.–E. Bursell
    Joslin Vision Network, Joslin Diabetes Center, Boston, MA
  • S.C. Nemeth
    Eye of the Wolf, LLC, Albuquerque, NM
  • Y. Srinivasan
    Biomedical, Kestrel, Albuquerque, NM
  • B. Raman
    Biomedical, Kestrel, Albuquerque, NM
  • Footnotes
    Commercial Relationships  G. Zamora, Kestrel Corporation, E; S. Bursell, None; S.C. Nemeth, Kestrel Corporation, C; Y. Srinivasan, Kestrel Corporation, E; B. Raman, Kestrel Corporation, E.
  • Footnotes
    Support  NIH Grant 1R43EY014493–01A1, DoD DAMD 17–03–2–0062
Investigative Ophthalmology & Visual Science May 2006, Vol.47, 974. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      G. Zamora, S.–E. Bursell, S.C. Nemeth, Y. Srinivasan, B. Raman; A Statistical Approach to Automatic Screening for Diabetic Retinopathy . Invest. Ophthalmol. Vis. Sci. 2006;47(13):974.

      Download citation file:

      © ARVO (1962-2015); The Authors (2016-present)

  • Supplements

Purpose: : To demonstrate the feasibility of a completely automatic screening system for Diabetic Retinopathy (DR) that employs a statistical approach to generate a diagnosis based on the presence of features in retinal images.

Methods: : A training set of images from 70 eyes (42 subjects) was obtained from Joslin Vision Network. Each eye set consisted of 45o stereo retinal color images from three different fields taken with a non mydriatic TopconTM fundus camera with 640 X 480 pixel resolution. For the ground truth scores, these images were manually graded by an ophthalmic technologist and a Joslin grader following the ETDRS protocol. An automatic image quality assessment tool was implemented based on the analysis of eight different textural features plus two vessel–related features. The DR screening system combines mathematical morphology, multivariate discriminant analysis and genetic algorithms (GA) to detect MAs. Techniques from mathematical morphology include: opening operations using linear rotating structuring elements to extract the blood vessels and morphological reconstruction, top–hat transform and Gaussian matched filtering to extract objects with MA characteristics. A set of 8 morphological features was computed for each detected MA, and a generalized linear model using the logit link function was trained on these to separate true positives from false positives. A GA was trained to combine the number of MAs detected in the three different imaging fields to make a final diagnosis about the presence or absence of DR.

Results: : The clinical sensitivity and specificity to detect DR automatically were 0.739 and 0.792, respectively with an accuracy of 0.757. The sensitivity and specificity of the DR diagnosis can be tuned by varying the field combination parameters in the genetic algorithm’s objective function.

Conclusions: : A system to automatically screening DR in digital retinal images of undilated eyes has been demonstrated. At the demonstrated levels of performance the system would allow a reading center to screen more patients with the same human resources and to extend its services to underserved populations through telemedicine infrastructure. Currently, the system is being integrated into Joslin’s operating environment for clinical testing and validation.

Keywords: diabetic retinopathy • imaging/image analysis: clinical 

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.