April 2009
Volume 50, Issue 13
Free
ARVO Annual Meeting Abstract  |   April 2009
Automated Detection of Regional Axonal Damage in Optic Nerve Cross-Section Images
Author Affiliations & Notes
  • J. Reynaud
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • G. Cull
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • L. Wang
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • B. Fortune
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • N. G. Strouthidis
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • C. F. Burgoyne
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • G. A. Cioffi
    Ophthalmology, Devers Eye Institute, Portland, Oregon
  • Footnotes
    Commercial Relationships  J. Reynaud, None; G. Cull, None; L. Wang, None; B. Fortune, None; N.G. Strouthidis, None; C.F. Burgoyne, None; G.A. Cioffi, None.
  • Footnotes
    Support  Merck & Co., Inc; Legacy Good Samaritan Foundation; NIH R01 EY11610 (CFB)
Investigative Ophthalmology & Visual Science April 2009, Vol.50, 5829. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      J. Reynaud, G. Cull, L. Wang, B. Fortune, N. G. Strouthidis, C. F. Burgoyne, G. A. Cioffi; Automated Detection of Regional Axonal Damage in Optic Nerve Cross-Section Images. Invest. Ophthalmol. Vis. Sci. 2009;50(13):5829.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose: : The first step in developing a method to count axons in damaged optic nerves (ONs) is to be able to detect whether an image fits a "normal" or "damaged" pattern. The purpose of this work is to develop an algorithm to perform this classification and an initial detection of regional ON damage.

Methods: : Histologic cross-sections from 2 normal (N1, N2) and 4 damaged (D1-D4) non-human primate ONs were imaged at 100X magnification. To compensate for illumination and stain differences, all images were normalized using their mean pixel intensity and standard deviation. To capture the textural differences between normal and damaged images, a compacted version of the Grey Level Co-ocurrence Matrix (cGLCM) was calculated for every image. 112 normal (from N1, N2) and 112 damaged (from D1-D3) images were selected and their cGLCMs used to train an Artificial Neural Network (ANN). The network performance was tested using 4687 normal and 3754 damaged images (from D3, D4) previously classified by a trained observer.

Results: : The ANN correctly classified 88.2% of normal and 95.7% of damaged images. Figures A and D are low power (5X) images of D3 and D4. The blue lines provide a first-order approximation of the normal/damage boundary. Figures B and E show the classification from a trained observer. Figures C and F show the ANN classification.

Conclusions: : The ANN accurately detects regions of axonal damage. The difference between normal and damaged classification rates is mainly due to the large number of normal images that are partially filled with axons (e.g. next to large vessels). A second ANN trained with damaged and partially filled normal images is being developed to help reduce this difference. By using the classification provided by the ANN, we plan to modify our existing algorithm to count viable axons present in regions of damage in ONs.

Keywords: optic nerve • image processing • imaging/image analysis: non-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.

×