DL Methods for Skin Lesions Automated Diagnosis In Smartphone Images

Skin lesions classification is a crucial endeavor in dermatology, as precise diagnosis is critical for improving patient outcomes and lowering the incidence of skin cancer. The PAD-UFES-20 dataset is very relevant for this task since it contains 2,298 imbalanced skin lesions smartphone images. Thus, data sampling techniques such as Random Over Sampling (ROS) and SMOTE were applied to overcome this issue. This work is based on two scenarios. The first scenario aims to classify the whole six classes figuring in this dataset, using various powerful proposed methods based on Deep Learning (DL) models. The best performance was achieved when employing EfficientNetB0 with data ROS, as it reached an accuracy of 92.01%, F1-score of 92.01%, and AUC of 97.12%. The objective of the second scenario is to make a triage between two skin lesions: Melanoma and Nevus using DL techniques. The results of the use of an EfiicientNetB1 with data ROS demonstrated a high level of accuracy, F1-score, and AUC, with a measurement of 98.98%, 99.22%, and 99.95%, respectively. The study’s high-performance shows that these methods might have the potential to be a useful tool for clinicians in the examination and diagnosis of skin lesions, thereby enhancing the healthcare quality.