April 2010
Volume 51, Issue 13
ARVO Annual Meeting Abstract  |   April 2010
Neural Network Algorithms for a Device to Measure Macular Visual Sensitivity
Author Affiliations & Notes
  • M. K. Smolek
    Ophthalmology, LSU Eye Center, New Orleans, Louisiana
  • K. Lebow
    Private Practice, Virginia Beach, Virginia
  • N. Notaroberto
    EyeCare 20/20, Slidell, Louisiana
  • A. Pallikaris
    Institute of Vision and Optics, University of Crete, Heraklion, Greece
  • S. Vujosevic
    Fondazione GB Bietti-IRCCS, Rome, Italy
  • Footnotes
    Commercial Relationships  M.K. Smolek, CenterVue, C; K. Lebow, CenterVue, I; CenterVue, C; CenterVue, R; N. Notaroberto, CenterVue, R; A. Pallikaris, CenterVue, R; S. Vujosevic, None.
  • Footnotes
    Support  Research to Prevent Blindness, Inc. (MKS); LSU Research Enhancement Fund (MKS)
Investigative Ophthalmology & Visual Science April 2010, Vol.51, 3550. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      M. K. Smolek, K. Lebow, N. Notaroberto, A. Pallikaris, S. Vujosevic; Neural Network Algorithms for a Device to Measure Macular Visual Sensitivity. Invest. Ophthalmol. Vis. Sci. 2010;51(13):3550.

      Download citation file:

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

  • Supplements

Purpose: : The MAIA macular integrity assessment device (CenterVue, SpA; Padova, Italy) measures visual sensitivity and fixation stability using 61 or 37 Goldmann-style stimulus points within a 10° meridional grid. Evaluation of average visual sensitivity in decibels (dB) for age-related macular degeneration (AMD) indicated a statistically significant difference compared to normal eyes. However, overlapping distributions reduced screening sensitivity, plus the use of an average stimulus value may hide early disease. The current study explored the use of artificial neural networks (NN) to evaluate individual MAIA stimulus values combined with summary statistics for the purpose of AMD screening.

Methods: : A total of 813 eyes (494 normal and 319 AMD) from subjects 21 to 92 years of age were collected at 5 test sites. Initial screening for AMD was performed by human experts using traditional clinical methods that did not include MAIA results. Randomly selected eyes were then used to build a NN training set composed of 20% of the data, while a test set was composed of 100% of the data. Two screening methods were compared. The Full Grid method used all 61 stimulus values, the average stimulus value, the number of points less than 24 dB (K-value), age, and the percentage of fixation points within a 1° and 2° area (P1 & P2) as NN inputs. The Sparse Grid method used 37 stimulus points (odd-numbered rings), average stimulus, K-value, and age as inputs. False Positive (FP), False Negative (FN), True Positive (TP) and True Negative (TN) results were tabulated based on the test set results. Suspect cases were defined as those with equivocal outputs (within 10% or 8% of being misgraded).

Results: : The Full Grid method produced 18 FP, 21 FN, 269 TP and 463 TN resulting in 92.8% sensitivity, 96.3% specificity and 5% suspects (28 AMD and 13 normals). The Sparse Grid method produced 7 FP, 34 FN, 271 TP and 484 TN resulting in 88.8% sensitivity, 98.6% specificity and 2% suspects (13 AMD cases and 3 normals).

Conclusions: : The Sparse Grid method had 4% reduced sensitivity compared to the Full Grid method, but higher specificity and fewer suspects. The Full Grid method had high sensitivity and specificity with the added assurance of more complete macular coverage, but required significantly longer test times. The Sparse Grid method may be best suited for annual AMD screening. The Full Grid method may be best suited for re-evaluating cases already diagnosed with AMD and who are being treated.

Keywords: age-related macular degeneration • clinical (human) or epidemiologic studies: systems/equipment/techniques • computational modeling 

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.