Automated melanoma detection: multispectral imaging and neural network approach for classification

Abstract

Our aim in the present research is to investigate the diagnostic performance of artificial neural networks (ANNs) applied to multispectral images of cutaneous pigmented skin lesions as well as to compare this approach to a standard traditional linear classification method, such as discriminant function analysis. This study involves a series of 534 patients with 573 cutaneous pigmented lesions (132 melanomas and 441 nonmelanoma lesions). Each lesion was analyzed by a telespectrophotometric system (TS) in vivo, before surgery. The system is able to acquire a set of 17 images at selected wavelengths from 400 to 1040 nm. For each wavelength, five lesion descriptors were extracted, related to the criteria of the ABCD (for asymmetry, border, color, and dimension) clinical guide for melanoma diagnosis. These variables were first reduced in dimension by the use of factor analysis techniques and then used as input data in an ANN. Multivariate discriminant analysis (MDA) was also performed on the same dataset. The whole dataset was split into two independent groups: i.e., train (the first 400 cases, 95 melanomas) and verification set (last 173 cases, 37 melanomas). Factor analysis was able to summarize the data structure into ten variables, accounting for at least 90% of the original parameters variance. After proper training, the ANN was able to classify the population with 80% sensitivity, 72% specificity, and 78% sensitivity, 76% specificity for the train and validation set, respectively. Following ROC analysis, area under curve (AUC) was 0.852 (train) and 0.847 (verify). Sensitivity and specificity values obtained by the standard discriminant analysis classifier resulted in a figure of 80% sensitivity, 60% specificity and 76% sensitivity, 57% specificity for the train and validation set, respectively. AUC for MDA was 0.810 and 0.764 for the train and verify set, respectively. Classification results were significantly different between the two methods both for diagnostic scores and model stability, which was worse for MDA.