April 2010
Volume 51, Issue 13
ARVO Annual Meeting Abstract  |   April 2010
Image Processing of Fundus Autofluorescence Images to Enhance Features
Author Affiliations & Notes
  • R. Lu
    School of Optometry, Indiana University at Bloomington, Bloomington, Indiana
  • B. P. Haggerty
    School of Optometry, Indiana University at Bloomington, Bloomington, Indiana
  • S. A. Burns
    School of Optometry, Indiana University at Bloomington, Bloomington, Indiana
  • A. E. Elsner
    School of Optometry, Indiana University at Bloomington, Bloomington, Indiana
  • Footnotes
    Commercial Relationships  R. Lu, None; B.P. Haggerty, None; S.A. Burns, None; A.E. Elsner, None.
  • Footnotes
    Support  NIH Grant EY007624
Investigative Ophthalmology & Visual Science April 2010, Vol.51, 2509. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      R. Lu, B. P. Haggerty, S. A. Burns, A. E. Elsner; Image Processing of Fundus Autofluorescence Images to Enhance Features. Invest. Ophthalmol. Vis. Sci. 2010;51(13):2509.

      Download citation file:

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

  • Supplements

Purpose: : To process multiple retinal images from fundus autofluorescence that are dim or have low signal to noise ratio, to maintain resolution and improve signal to noise level while maintaining the illumination at safe levels. To use subpixel alignment techniques to improve the clarity or detectability of fine features such as small retinal blood vessels, hypofluorescence, or hyperfluorescence, as well as other fundus structures.

Methods: : Fundus autofluorescence images acquired with the research Scanning Laser Opthalmoscope with 594 nm excitation at 92 microwatts require processing of multiple images, but do delineate features. To obtain images at these safe light levels, the detector gain must be so high that an unusual amount of noise is present in each image, and the signal is still low. However, the optical resolution exceeds our digital resolution, and there are large numbers of images (92) captured in a short time. The field size is sufficient (27 x 23 deg) to provide robust features for alignment, such as larger retinal vessels and the optic nerve head. Thus, we investigated a family of related techniques to further improve image quality, using customized MATLAB routines. We compared upsampling images by 2X in two dimensions vs. no upsampling. The upsampling adds periodic noise, and we varied the content of the added matrix elements and methods of smoothing, prior to performing alignment by cross correlation. For the added matrix elements, we used 0, the local mean, a replication of the adjacent line, smoothing with a Gaussian or median filter, and the Lanczos2 algorithm kernel operating on the adjacent pixel values. We compared aligning individual images vs. aligning small groups of averaged images. We assessed images for improvement: better visibility of small retinal vessels, detecting small hyperfluorescent or hypofluorescent features, and visualizing retinal pigment epithelial cells.

Results: : The improvement from upsampling was most noticeable for small blood vessels, which block autofluorescence. Additional branches of small retinal vessels became visible in the upsampled images. The blood vessels had a more uniform gray appearance and appeared continuous. Patchy autofluorescence was also clearer in some areas.

Conclusions: : Upsampling, when possible, can better delineate some of the small features in fundus autoflurescence, without the need for increased light.

Keywords: image processing • retinal pigment epithelium 

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.