May 2005
Volume 46, Issue 13
ARVO Annual Meeting Abstract  |   May 2005
Automated Pattern Recognition for the Detection Diabetic Changes in Digital Fundus Images
Author Affiliations & Notes
  • J. Forsstrom
    Mikkeli Polytechnic, Mikkeli, Finland
  • V. Kalesnykiene
    University of Kuopio, Kuopio, Finland
  • M. Kuivalainen
    Lappeenranta University of Technology, Lappeenranta, Finland
  • I. Sorri
    University of Kuopio, Kuopio, Finland
  • H. Uusitalo
    University of Kuopio, Kuopio, Finland
  • J. Kämäräinen
    Lappeenranta University of Technology, Lappeenranta, Finland
  • Footnotes
    Commercial Relationships  J. Forsstrom, None; V. Kalesnykiene, None; M. Kuivalainen, None; I. Sorri, None; H. Uusitalo, None; J. Kämäräinen, None.
  • Footnotes
    Support  None.
Investigative Ophthalmology & Visual Science May 2005, Vol.46, 3958. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      J. Forsstrom, V. Kalesnykiene, M. Kuivalainen, I. Sorri, H. Uusitalo, J. Kämäräinen; Automated Pattern Recognition for the Detection Diabetic Changes in Digital Fundus Images . Invest. Ophthalmol. Vis. Sci. 2005;46(13):3958.

      Download citation file:

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

  • Supplements

Abstract: : Purpose: To automatically classify fundus images as suspected diabetic retinopathy or normal, in order to decrease the amount of fundus images to be reviewed by a specialist in broad–scale screening for diabetic retinopathy Methods: Pattern recognitions algorithms were developed for off–line analysis of the fundus image. Uncompressed 50 degrees colored tiff–images with 1500x1150 resolution were used. The algorithms recognize different lesions (red small dots, i.e. microaneurysms and very small hemorrhages, normal hemorrhages and soft and hard exudates) as well as blood vessels, papilla and light reflection artefacts. A quality control algorithm was implemented to reject bad quality images. For lesion detection, brightness normalization was used. Vessel detection was based on structural analysis, being robust against light conditions. The algorithms were tested using a test set of 45 images, where 20 were lesion free. The images were annotated by a specialist. Results: For screening purposes, high sensitivy is important, while the specifity and ability to analyze images of poor quality is less critical. It has been important to combine the lesion detection with vessel and artefact detection to increase the specifity, which in turn allows adjusting the algorithms for higher sensitivity. The measured sensitivity for hard and soft exudates, together, was 90.5% (in the test set 21 images of 45 contained exudates), 100% for hemorrhages (existed in 26 images) and 96% for red small dots (existed in 25 images). The specifity was 75% for exudates, 65% for hemorrhages and 65% for red small dots. It is noted that combining the findings of different lesions will increase sensitivity and decrease specifity. With the given results, an adjustment towards higher specifity is possible. Conclusions: The results indicate that the developed algorithms and combination of them have potential to become a first phase screening tool, saving work from specialists in broad–scale screening.

Keywords: diabetic retinopathy • detection • retina 

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.