Comparative Evaluation of Advanced Deep Learning Approaches for Melanoma Detection in Dermoscopic Images Toward Supporting Clinical Diagnosis

Abstract

Melanoma is a type of skin cancer that is one of the most aggressive, and its diagnosis requires quick and accurate identification to increase patient survival. This paper draws a comparative analysis of advanced deep learning systems, namely, VGG and traditional CNN systems, used in automatic melanoma detection in dermoscopic images. The techniques included preprocessing of dermoscopic sample datasets, training of both VGG and baseline CNN, and the testing of their classification by confusion matrices and receiver operating characteristic (ROC) analysis. The results showed that VGG model was more inclined to have high true positive rate of melanoma, and had 20/20 correct melanoma and 0.92 area under ROC curve (AUC) in its confusion chart. The CNN model did a little better, with a slightly higher AUC of 0.95 and lower false negative percentages at normal. These results show that deep learning models can be used to advance clinical and face-to-face diagnosis of melanoma; both models provide strong performance.