June 2015
Volume 56, Issue 7
ARVO Annual Meeting Abstract  |   June 2015
An algorithm combining structure and function matches the diagnoses of glaucoma specialists with high accuracy
Author Affiliations & Notes
  • Hassan Muhammad
    Psychology, Columbia University, New York, NY
  • Ali S Raza
    Psychology, Columbia University, New York, NY
  • Donald Hood
    Psychology, Columbia University, New York, NY
  • Footnotes
    Commercial Relationships Hassan Muhammad, None; Ali Raza, None; Donald Hood, Topcon (F)
  • Footnotes
    Support None
Investigative Ophthalmology & Visual Science June 2015, Vol.56, 630. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Hassan Muhammad, Ali S Raza, Donald Hood; An algorithm combining structure and function matches the diagnoses of glaucoma specialists with high accuracy. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):630.

      Download citation file:

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

  • Supplements

Purpose: To assess the diagnostic ability of a combined structure and function, cluster method,[1] by comparing its results to those of glaucoma specialists.

Methods: One eye of 46 patients was tested with 24-2 visual fields (VF) and frequency-domain optical coherence tomography (fdOCT). Four glaucoma specialists judged that 30 of the 46 patient eyes had glaucomatous damage, and 16 were “healthy”, forming our “reference standard”. These eyes were part of a study of 50 eyes with an abnormal appearing disc and a 24-2 mean deviation better than -6dB.[2] After manually centering the fovea and correcting algorithm segmentation of macular fdOCT cube scans, retinal ganglion cell plus inner plexiform layer (RGC+) thickness maps were generated, down-sampled to 16x16 grids, and converted to probability maps based upon a normative group of 51 healthy controls.[2] The largest cluster was found (using p=0.05 as a threshold) within a hemi-retina/field with points in the center and edges omitted.[1] A minimum probability value was determined and adjusted based on Monte-Carlo simulations. This clustering method was repeated on the 24-2 VF total deviation values without omitting points, and the results were combined with the fdOCT grid to produce a continuous metric, PCC [1]. To test this metric, the results were compared to the reference standard and false positives (FP) and false negatives (FN) were determined.

Results: The area-under-the-curve was 0.90, indicating a strong discrimination between the abnormal and healthy eyes. The following sensitivities (SN) and specificities (SP) were observed: 73% SN at 96% SP (4 FP/1 FN) and 87% SN at 87% SP (2 FP/4 FN). Three of 4 FPs showed poor local agreement between VFs and fdOCT, although there was agreement based upon the overall hemi-retina/field comparison used. Further, the RGC+ plots had circum-foveal abnormalities associated with normal anatomical differences. The FNs tended to have subtle damage in an arcuate pattern.

Conclusions: This cluster method showed good agreement with the glaucoma specialists, correctly identifying 89% of the eyes with a specificity of 96%. It should be possible to improve this accuracy by considering the relative positions of OCT and VF clusters within a hemi-retina/field, as well as excluding a larger area from the center of the fdOCT grid. 1. Raza, AS, et al., 2014, IOVS; 2. Hood et al., 2014, TVST.


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.