April 2011
Volume 52, Issue 14
ARVO Annual Meeting Abstract  |   April 2011
Automized 3d Angiography With Doppler Oct Using A Support Vector Machine
Author Affiliations & Notes
  • Amardeep S. Singh
    Center for Medical Physics, Medical University Vienna, Vienna, Austria
  • Tilman Schmoll
    Center for Medical Physics, Medical University Vienna, Vienna, Austria
  • Rainer A. Leitgeb
    Center for Medical Physics, Medical University Vienna, Vienna, Austria
  • Footnotes
    Commercial Relationships  Amardeep S. Singh, None; Tilman Schmoll, None; Rainer A. Leitgeb, None
  • Footnotes
    Support  European FP 7 HEALTH program (grant 201880, FUN OCT).
Investigative Ophthalmology & Visual Science April 2011, Vol.52, 1711. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Amardeep S. Singh, Tilman Schmoll, Rainer A. Leitgeb; Automized 3d Angiography With Doppler Oct Using A Support Vector Machine. Invest. Ophthalmol. Vis. Sci. 2011;52(14):1711.

      Download citation file:

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

  • Supplements

Purpose: : To develop a method that is capable of reliably extracting Doppler OCT (DOCT) signatures with minimal user intervention and that is more robust to multiple scattering and phase noise than simple threshold based methods. The algorithm is applied to extracting dynamic perfusion parameters in-vivo from circumpapillary scans and to provide 3D angiography maps.

Methods: : We employ a support vector machine (SVM) in order to combine different features of a pixel in a Doppler OCT image, that contrasts perfusion from static tissue areas. We used three features: the intensity, average intensity in a 2x2 window and the entropy in a 5x5 window. For the entropy the DOCT image is masked with the intensity image in order to only have relevant information in the image present. This narrows the histogram and by using histogram equalization, the intensities of the flow are mapped to nearly constant values, whereas the bulk tissue becomes very grainy. This allows for the application of the entropy filter for differentiation of tissue and vessels. The features are combined by training a 1-norm SVM using a radial-basis function kernel. After training it once, it can be readily applied in order to classify new DOCT data sets. Doppler data have been measured on healthy subjects with a CMOS based fast DOCT platform operating from 30kHz to 200kHz.

Results: : We applied the procedure to in-vivo circular DOCT scans around the optic nerve head and to a volume scan of a healthy subject. The extracted quantitative angiograms are compared to those obtained with a Gaussian mixture model (GMM) that was trained on the intensity histogram of the same data that was used for training of the SVM. The GMM which is a threshold based procedure also segmented some of the phase noise and the multiple scattering. Our proposed algorithm is more robust to multiple scattering as it takes into account spatial information around the pixels. Moreover it does not need manual intervention after training in order to find optimal segmentation results.

Conclusions: : We apply the method to circular scans and a volume scan of a healthy volunteer and show that our method consistently performs better than other standard segmentation methods. It might be envisioned to include more features in the analysis although the danger of overfitting the SVM exists. We believe this result to be an important step towards complete automized analysis of DOCT data.

Keywords: image processing • blood supply • 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.