April 2010
Volume 51, Issue 13
ARVO Annual Meeting Abstract  |   April 2010
Improving the Classification of Abnormal Optic Discs Using Shape Analysis
Author Affiliations & Notes
  • H. Zhu
    Optometry and Visual Science, City University, London, United Kingdom
  • D. P. Crabb
    Optometry and Visual Science, City University, London, United Kingdom
  • P. H. Artes
    Ophthalmology and Visual Sciences, Dalhousie University, Halifax, Nova Scotia, Canada
  • Footnotes
    Commercial Relationships  H. Zhu, None; D.P. Crabb, None; P.H. Artes, None.
  • Footnotes
    Support  None.
Investigative Ophthalmology & Visual Science April 2010, Vol.51, 2727. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      H. Zhu, D. P. Crabb, P. H. Artes; Improving the Classification of Abnormal Optic Discs Using Shape Analysis. Invest. Ophthalmol. Vis. Sci. 2010;51(13):2727.

      Download citation file:

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

  • Supplements

Purpose: : To develop a shape analysis model to extract shape features from optic disc topography acquired by Heidelberg Retina Tomography (HRT) and to use this information to improve classification of abnormality.

Methods: : Mean HRT topographies of 164 healthy subjects, 252 glaucomatous patients (as described in Coops et al 2006) and 666 suspect glaucomatous eyes were investigated. Instead of using area and volume parameters from the HRT software, all topography images were processed with a shape analysis model (a non-linear latent variable Gaussian process model): this extracted multivariate shape features from the topographies. A feature map derived from this model placed each optic disc in a unique position in a multivariate space. Discs were then classified using a criterion of at least one sector being abnormal or on borderline by HRT Moorfields Regression Analysis (MRA). For the eyes with false positive (FP) and false negative (FN) classifications, two multivariate Gaussian distribution in the feature map were fitted respectively and were considered as the least specific region (LSpeR) and the least sensitive region (LSenR).

Results: : The shape model required no diagnostic information or manual input and yielded objective shape features of optic disc topography. The three most important features appeared to relate to disc size, cup depth and disc tilt, but more than 100 ‘features’ were extracted from the disc topography. Sensitivity and specificity for MRA in the sample of subjects with diagnosis was 81% and 69% respectively. Forty-eight (including 45 FNs with MRA classifications) and 55 (including 50 FPs with MRA classifications) discs were bounded by LSenR and LSpeR respectively: only 3 non-FN discs (including 2 true-positives and 1 true-negative) and 5 non-FP discs (including 4 true-negatives and 1 true-positive) fell in LSpeR and LSenR respectively. LSpeR and LSenR gave a good indication of those eyes in which FP and FN errors would be expected with the MRA.

Conclusions: : The feature map formed by the shape model identified previously unused information about individual optic disc topography. The map provided ‘restricted regions' where the diagnostic performance of conventional analysis methods was limited. The subjects with topography shape features in these regions should be referred to other methods of glaucoma diagnosis. Conversely, subjects bounded in other areas of the feature map can be classified with greater certainty. The clinical utility of the technique needs to be established in other datasets.

Keywords: imaging/image analysis: clinical • optic disc 

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.