May 2008
Volume 49, Issue 13
Free
ARVO Annual Meeting Abstract  |   May 2008
A Multilayerd Approach to the Automatic Analysis of mfEERG Waveforms: The Artificial Neural Network
Author Affiliations & Notes
  • A. A. Foulis
    Electrodiagnostic Imaging Unit Ophthalmology, Gartnavel General Hospital, Glasgow, United Kingdom
  • S. Parks
    Electrodiagnostic Imaging Unit Ophthalmology, Gartnavel General Hospital, Glasgow, United Kingdom
  • D. Keating
    Electrodiagnostic Imaging Unit Ophthalmology, Gartnavel General Hospital, Glasgow, United Kingdom
  • Footnotes
    Commercial Relationships  A.A. Foulis, None; S. Parks, None; D. Keating, None.
  • Footnotes
    Support  None.
Investigative Ophthalmology & Visual Science May 2008, Vol.49, 2217. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      A. A. Foulis, S. Parks, D. Keating; A Multilayerd Approach to the Automatic Analysis of mfEERG Waveforms: The Artificial Neural Network. Invest. Ophthalmol. Vis. Sci. 2008;49(13):2217.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose: : The multifocal ERG (mfERG) provides spatial and temporal information on the retina’s function in an objective manner making it a valuable tool for diagnosing retinal abnormalities, however interpretation of the signals is often difficult. A system capable of automatically classifying the waveforms and comparing serial visits would therefore be advantageous.

Methods: : A multilayered approach is being studied to achieve this goal. A range of information will be included such as, analysis of the Fourier domain profiles, wavelet transforms, filters, signal to noise ratio mapping, amplitude and latency information and the use of artificial neural networks (ANN). Waveforms will be studied both individually and in a spatial context. This study describes the application of ANNs to classify the mfERG.Waves were to be categorized as either normal, reduced in amplitude, delayed, reduced and delayed or as having no significant response. 40 clinical waveforms were selected as base waves with which to create the training set using a series of manipulations. Both periodic and random noise were added in varying degrees, as was drift to simulate factors that degrade the quality of the signal and hence increase the difficulty of interpretation. 4300 of these waves were selected for the training set, while 1000 were used to form a validation set. A further testing set was constructed using 1000 true waves illustrating various conditions and recording qualities. The performance of a network was assessed by comparing its classification for a waveform with that stated by an experienced operator. The validation set was used for the initial testing phase and if the network performed well, the testing set was used. Parameters such as the network architecture, the number of neurons in the hidden layer, the neuron transfer function, the number of training iterations and the learning rate were varied until an optimal performance was achieved.

Results: : A feed-forward back-propagation network with one hidden layer was found to yield the optimal performance. 93% of those in the validation set and 70% of those in the testing set were classified correctly. Waves close to the timing and amplitude boundaries were the most problematic for the network.

Conclusions: : ANNs are a viable component in a multilayered approach to simplify the interpretation and analysis of the mfERG waveforms.

Keywords: electroretinography: non-clinical • 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.

×