Article Text

Download PDFPDF

Progressive assessment of age related macular degeneration using an artificial neural network approach
  2. N K TAYLOR,
  1. Medical Imaging RG, Geomatics Unit, Faculty of Environmental Studies, ECA/Heriot Watt, 79 Grassmarket, Edinburgh EH1 2HJ, UK
  1. justin{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Editor,—The key to successful age related macular degeneration (ARMD) screening is the efficient production of accurate classified images with minimum patient inconvenience.1 The technologies of digital image analysis and artificial neural networks (ANN) are not new and have been used in the past to provide a more objective basis for a range of medical applications.2-9They have, however, not been used for operational classification of maculopathies such as ARMD. Research has shown that ANN computer programs are capable of discriminating normal fundus from those with diabetic retinopathy, potentially reducing the numbers of images requiring expert examination by 70% or more.10

Digital fundus images from a Topcon Imagenet camera were modified by in-house computer imaging techniques (erdas Imagine Software) within a geographical information system (GIS) (Fig 1). The ANN used was a standard back propagation multilayer perceptron, running within the Stuttgart neural network system (snns) program.

Figure 1

Flow diagram of image analysis methodology.

The validating criteria for the study came from a small group of ophthalmologists carrying out masked assessment of a series of fundus images (stereo 35 mm slides and digital fundus images) which either contained ARMD at different confidence levels of judgment or which did not contain ARMD.


A total of 87 sample images of the drusen features under analysis were extracted from the postprocessed fundus images, of which 30 were used as test samples for operational accuracy assessments. Each extracted image represented a square sample of 11 × 11 (121) pixels of intensity information (the intensity information was the relative “brightness” of the feature within the range, 1–255 black to white). The feature types assessed constituted three subcategories within the grading system hierarchy; these were hard drusen (HD), large soft distinct drusen (LSD), and serogranular drusen (SGD). All other features indicative of ARMD such as haemorrhages, blood vessels, etc, were grouped into the background fundus class (FB) actively excluding them from further assessment. The choice of kernel size at 11 × 11 pixels was chosen because the main focus at this stage of the research was to differentiate the drusen subtypes using hard distinct drusen as the main discriminator and generally all sample features fit well within this window size. The ANN required approximately 10 000 iterations in order to categorise the training samples to within the accepted error margin of 0.01 and, after being analysed by both the expert assessor and the ANN model, the sample outputs were finally compared as shown in Table 1 to assess the accuracy of the computer based analysis system with the clinical standard.

Table 1

Error matrix of ANN assessment against reference (clinical) assessment

The ANN, compared with the reference assessment acrossall test feature classes, yielded an overall accuracy of 69.21%, with sensitivity to drusen classes (HD/SGD) being 95%/99% respectively and specificity 55% and 75%.

The overall accuracy of the ANN test method across thetwo test feature classes was found to be 66%, with 72% and 90% sensitivity (HD/SGD) and 72% and 63% Specificity respectively. Current published literature on ANN pattern recognition tasks suggests that results of ∼70% overall accuracy indicate a good result for first stage ANN analyses. The results obtained in this study with values of (95%+ sensitivity, 75% specificity) indicate that both types of drusen are being clearly differentiated by the ANN

The neural network was trained to an accuracy of within 0.01 for each drusen subtype (hard, serogranular, and large soft distinct drusen) before the validation set was classified. Results indicate that the neural network performed better with more numerous feature classes available; the system sensitivity overall being found to be 95% with 75% specificity.


A simple methodology for using computer based image processing and feature detection techniques to accurately quantify drusen has been presented and results are comparable with clinical trials. This approach could be applied to operational assessment of fundus diseases providing benefits both in time management and associated cost.


Thanks are due to the Gift of Thomas Pocklington and the Royal College of Surgeons for funded support towards this study.