Computer-aided detection of melanoma a case study

Abstract

Melanoma is the most dangerous form of skin cancer, and its detection at an early stage can allow timely treatment and prevention of fatal consequences. In this paper we present a case study of computer-aided diagnostics of melanoma using images of patients moles. The initial study was performed on two datasets: a benchmark dataset which is publicly available and a second one, containing images that were taken in hospitals in Macedonia. We present the obtained results and a short discussion of further directions for research. The results on the initial dataset were promising and showed 83% accuracy in the detection of the melanoma on the benchmark dataset. However, the same approach applied on the Macedonian dataset, the results could not be reproduced due to the low number of positive examples. The results showed that the performance of the classifiers did not benefit from under-sampling or oversampling techniques, nor did from feature selection. We can conclude that to build a reliable system for melanoma detection, a datasets of hundreds of images is not enough to train a machine-learning based model.