Improving binary skin cancer classification based on best model selection method combined with optimizing full connected layers of Deep CNN

Melanoma is one of the most dangerous skin cancers, leading to high mortality rates. Early detection and resection are two important steps to reduce mortality. Recently, several studies have applied artificial intelligence to facilitate binary skin cancer classification. However, the imbalance of the sensitivity and specificity metrics affects the overall performance of the models. To solve this problem, we propose an optimization for deep Convolutional Neural Network (CNN) combined with a change in the best model selection for this binary melanoma classification. Our research uses ISIC 2019, the latest and largest dataset consist of 17,302 skin lesion images for training and best model selection. The performance of the best models is analyzed based on the 10% data of ISIC 2019 dataset, then compared with the performance of dermatologists on 100 medical images of the MClass-D dataset. Our optimized deep CNN solves the underfitting problem and avoids overfitting. Our proposed best model selection method with an increase in Youden Index (YI) on both test-10 and MClass-D datasets also outperforms traditional methods. Moreover, our solution effectively outperformed 153 out of 157 dermatologists, which surpasses the current state-of-the-art solution by 17 dermatologists.