Segmentation of virus-infected areas in retinal angiograms using a learning-by-sample approach

A operational system devoted to the segmentation of virus-infected areas in the retina is described. It uses a 3-stage approach which involves image sampling, unsupervised coding and supervised classification. Unsupervised coding is provided by principal component analysis whereas supervised classification is performed by a multilayer perceptron. Segmentation as realized by ophthalmologists is considered to be the gold standard. It is shown that, despite the high variability of images, automatic segmentation is accurate and can help to spot problematic areas.