Application of an artificial neural network in epiluminescence microscopy pattern analysis of pigmented skin lesions: a pilot study.

In vivo epiluminescence microscopy (ELM) is a non-invasive technique which improves the clinical diagnosis of naevi and malignant melanoma by providing diagnostic criteria that cannot be appreciated by the naked eye. The present study investigated whether ELM criteria pattern analysis can be employed in an objective, observer-trained, computer-aided diagnostic system, and whether artificial neural networks (ANN) can be applied to the diagnosis of pigmented skin lesions (PSL). The ELM criteria patterns of 200 PSL oil immersion images (60 common naevi, 60 dysplastic naevi, and 80 malignant melanomas) were analysed using a standardized questionnaire. One hundred randomly assigned PSL were used as a training set for an ANN, the remaining 100 PSL serving as the test set. The ANN was trained by backward propagation according to the histological diagnosis, and its performance was compared with that of human investigators. Out of the test set the human investigators correctly diagnosed 88% of PSL and the ANN 86%. In a dichotomized model comparing common, compound, and dysplastic naevi vs. malignant melanoma, i.e. benign vs. malignant PSL, the sensitivity and specificity of human diagnosis was 95 and 90%, respectively, whereas the sensitivity and specificity of the ANN diagnosis was 95 and 88%. Our data indicate that artificial neural networks can be trained to diagnose PSL at a human expert level, based on patterns provided by ELM criteria. We suggest that this technique offers a new approach to the diagnosis of PSL.