Shallow and deep learning approaches to classify melanoma and non-melanocytic skin lesions

Several image processing methods in Dermatology are grounded in shallow and deep learning approaches. These solutions are relevant to assist health experts in decision-making processes related to harmful melanoma-a malignant melanocytic condition-and other skin lesions. This work aims to compare these approaches in a specific classification problem: malignant melanocytic lesions versus non-melanocytic ones. We developed 39 learning method configurations, including three original ones based on fine-tuned deep neural networks. Some implemented settings incorporate auxiliary procedures, such as oversampling, feature selection and data augmentation. An experimental evaluation in the public Derm7pt dermoscopic database suggests that the best original setting performance was competitive against the leading results reported by recent literature alternatives. In particular, the proposal reached average accuracy and sensitivity of 0.9909 and 0.9976, respectively. These results were averaged across three runs of the stratified nested cross-validation strategy. Moreover, our 39 configurations outperformed an experimental baseline derived from the majority class error. Thus, this work can be helpful in inspiring computational systems that could act as preliminary filters to support the detection of a harmful form of skin cancer and its separation from other lesions.