April 2010
Volume 51, Issue 13
Free
ARVO Annual Meeting Abstract  |   April 2010
Automated Retinal and NFL Segmentation in OCT Volume Scans by Pixel Classification
Author Affiliations & Notes
  • K. A. Vermeer
    Rotterdam Ophthalmic Institute,
    Rotterdam Eye Hospital, Rotterdam, The Netherlands
    i-Optics Nederland BV, Rijswijk, The Netherlands
  • J. van der Schoot
    Rotterdam Ophthalmic Institute,
    Glaucoma Service,
    Rotterdam Eye Hospital, Rotterdam, The Netherlands
  • J. F. De Boer
    Rotterdam Ophthalmic Institute,
    Rotterdam Eye Hospital, Rotterdam, The Netherlands
    Dept. of Physics and Astronomy, VU University, Amsterdam, The Netherlands
  • H. G. Lemij
    Glaucoma Service,
    Rotterdam Eye Hospital, Rotterdam, The Netherlands
  • Footnotes
    Commercial Relationships  K.A. Vermeer, None; J. van der Schoot, None; J.F. De Boer, OCT technology, P; H.G. Lemij, None.
  • Footnotes
    Support  None.
Investigative Ophthalmology & Visual Science April 2010, Vol.51, 219. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      K. A. Vermeer, J. van der Schoot, J. F. De Boer, H. G. Lemij; Automated Retinal and NFL Segmentation in OCT Volume Scans by Pixel Classification. Invest. Ophthalmol. Vis. Sci. 2010;51(13):219.

      Download citation file:


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

      ×
  • Supplements
Abstract
 
Purpose:
 

To develop and evaluate an automated method for segmentation of the retina and the retinal nerve fiber layer (RNFL) in OCT volume scans.

 
Methods:
 

Volumetric scans of 10 normal and 8 glaucomatous subjects were acquired with a Spectralis OCT (Heidelberg Engineering) (20°x20° centered at the optic disk, 193 B-scans).Two B-scans of each healthy subject and one B-scan of each glaucomatous subject were manually segmented into four areas, defining three interfaces: vitreous/RNFL, RNFL/GCL and RPE/choroid.Each pixel along each A-line was augmented by simple averaged data from neighboring pixels on that same A-line, producing a feature vector for each pixel in the volumetric scan. The manually segmented scans of the healthy subjects were used to train a second degree polynomial support vector machine. The resulting classifier was then applied to the full volumetric scans and the results were regularized by a level set segmentation.Local accuracy was estimated by the root mean square of the differences between manually and automatically determined interfaces. For the data from healthy subjects, cross-validation was used.

 
Results:
 

Examples of the RNFL thickness of a (left) healthy and a (right) glaucomatous eye are shown below. For the healthy subjects, the estimated RMS errors of the interfaces were 3.3 µm (vitreous/RNFL), 4.2 µm (RNFL/GCL), 12.8 µm (RPE/choroid), excluding errors due to artifacts near the optic disk or at the boundaries of the scans. For glaucomatous eyes, the errors were 4.6 µm, 5.6 µm and 13.2 µm respectively.Worst performance was encountered in eyes with almost indiscernible contrast between NFL and GCL reflectivity.  

 
Conclusions:
 

The proposed method provides accurate interface localization for the vitreous/NFL and RPE/choroid interfaces. The NFL/GCL interface sometimes locks to other structures, yielding somewhat less accurate results.The resulting thickness maps match the expected morphological structure of the retina and the RNFL.

 
Keywords: image processing • retina • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound) 
×
×

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.

×