A Comprehensive Evaluation and Benchmarking of Convolutional Neural Networks for Melanoma Diagnosis

Melanoma is the most invasive skin cancer with the highest risk of death. While it is a serious skin cancer, it is highly curable if detected early. Melanoma diagnosis is difficult, even for experienced dermatologists, due to the wide range of morphologies in skin lesions. Given the rapid development of deep learning algorithms for melanoma diagnosis, it is crucial to validate and benchmark these models, which is the main challenge of this work. This research presents a new benchmarking and selection approach based on the multi-criteria analysis method (MCDM), which integrates entropy and the preference ranking organization method for enrichment of evaluations (PROMETHEE) methods. The experimental study is carried out in four phases. Firstly, 19 convolution neural networks (CNNs) are trained and evaluated on a public dataset of 991 dermoscopic images. Secondly, to obtain the decision matrix, 10 criteria, including accuracy, classification error, precision, sensitivity, specificity, F1-score, false-positive rate, false-negative rate, Matthews correlation coefficient (MCC), and the number of parameters are established. Third, entropy and PROMETHEE methods are integrated to determine the weights of criteria and rank the models. Fourth, the proposed benchmarking framework is validated using the VIKOR method. The obtained results reveal that the ResNet101 model is selected as the optimal diagnosis model for melanoma in our case study data. Thus, the presented benchmarking framework is proven to be useful at exposing the optimal melanoma diagnosis model targeting to ease the selection process of the proper convolutional neural network architecture.