Glaucoma Detection Using An Artificial Neural Network

1
Glaucoma Detection Using An Artificial Neural
Network

• SPS Grewal
• Dilraj Grewal
• Rajeev Jain
• V Rihani

Grewal Eye Institute, Chandigarh, India
Punjab Engineering College, Chandigarh, India
drgrewal_at_gmail.com
2
Introduction
Glaucoma A Diagnostic Dilemma

• There are a number of tools available for
  glaucoma with different diagnostic approaches.
  However the sensitivity and specificity of any
  single diagnostic test is not adequate to be
  considered the gold standard. Combining the
  information generated from different instruments
  can theoretically have better diagnostic
  capability.
• Methods of early detection of glaucoma often
  focus on Optic Disc Topography and RNFL
  thickness, besides IOP to identify patients
  likely to develop visual field defects.
• In an attempt to classify the eye as normal or
  glaucomatous, analysis strategies involve
  analysis of input parameters from different
  instruments using linear discriminant function
  (LDF) or Artificial neural networks (ANN) or an
  experienced Glaucoma specialist.
• We aim to classify the eyes as normal or
  glaucomatous based on input parameters from
  different instruments using an Artificial Neural
  Network. Multilayer perceptrons (MLP) with back
  propagated learning were used.

3
Methods
Artificial Neural Network

• A Multilayer Perceptron model (MLP) with back
  propagated learning method based Artificial
  Neural Network (ANN) for glaucoma detection was
  trained and tested using EasyNN-plus (v6.0g)
  software simulator.
• The modeled ANN had 114 input nodes corresponding
  to each parameter from the tests conducted
  Perimetry (Octopus 1-2-3), RNFL analysis on
  Scanning Laser Polarimetry (GDx Nerve Fibre Layer
  Analyzer Carl Zeiss Meditec, Inc., Dublin CA)
  and RNFL and ONH analysis on OCT (Carl Zeiss
  Meditec Inc., Dublin, CA) as well as patient
  history. There was 1 output node for classifying
  the eye as normal or glaucomatous.
• The ANN model had 3 hidden layers

4
Methods
Observerational, Retrospective, Non-randomized
Study
80 eyes were classified using optic disc and
perimetry into
Input Parameter
Collect Patient Records
Input Layer had 114 nodes and included data for
Preprocess Data
Build network with 114 input nodes and 1 output
node
Train the Network
Optimize Network
The data was analyzed for accuracy with two and
three outputs. The classification model was
cross validated using 20 eyes (normals5 POAG
suspects8 POAG7).
5
Results
Mean Age of patients 54.4 13.8 years
Average IOP 21.6 7.6 mmHg
Smax/Imax ROC curve
Cup Area (OCT) ROC curve
Smax Imax ROC curve
6
Discussion Conclusion

• Based on the 114 input parameters the artificial
  neural network assigned the maximum significance
  to the RNFL OCT parameters followed by the Optic
  Nerve Head OCT parameters.
• The neural network was trained using a supervised
  learning by a set of inputoutput pairs. The
  input was the raw data from the visual fields,
  GDx and OCT and the output was the classification
  made by the neural network. During training, the
  network was told whether it is correct or not,
  based on a gold standard (an experienced glaucoma
  specialist), and, after each session, it adjusted
  its internal parameters to increase the accuracy.
  The process was repeated till networks accuracy
  did not improve.
• The area under the ROC curve was the maximum for
  the inferior quadrant thickness on OCT (0.755)
  which compares well with studies conducted by
  Medeiros et al.(0.92),1 and Huang Chen. (0.83)2

• An Artificial Intelligence, ANN classification
  model, with data inputs from Perimetry GDx and
  OCT, was able to separate
  glaucomatous eyes from normals with a high degree
  of accuracy.
• Accuracy however dropped on increasing outputs
  to 3 as the network had a tendency to mislabel
  POAG suspects as POAG.
• The human based, evaluator bias is eliminated.
• The trainable ANN model considers all
  available parameters to give a consolidated,
  unbiased conclusion in glaucoma cases.
• Such automated classifier systems have good
  potential in diagnosing multi- faceted diseases
  like glaucoma. However the clinical acumen
  of an experienced glaucoma specialist still
  remains the gold standard.
• ANNs can be a useful tool with the general
  ophthalmologist.

1Medeiros FA, Zangwill LM, Bowd C, et al.
Comparison of the GDx VCC scanning laser
polarimeter, HRT II confocal scanning laser
ophthalmoscope, and Stratus OCT optical coherence
tomography for the detection of glaucoma. Arch
Ophthalmol. 2004122827837 2Huang ML, Chen HY.
Development and comparison of automated
classifiers for glaucoma diagnosis using Stratus
optical coherence tomography. Invest Ophthalmol
Vis Sci. 2005 Nov46(11)4121-9.