End-to-End Decoupled Training: A Robust Deep Learning Method for Long-Tailed Classification of Dermoscopic Images for Skin Lesion Classification

Due to its increasing incidence, skin cancer, and especially melanoma, is considered a major public health issue. Manually detecting lesions (SL) from dermoscopy images difficult time-consuming process. Thus, researchers designed computer-aided diagnosis (CAD) systems assist dermatologists in the early detection of cancer. Moreover, SL naturally exhibits long-tailed distribution due complex patient-level conditions existence rare diseases. Very limited research for handling this issue exists on detection. In paper, we propose an end-to-end decoupled training lesion classification task. Specifically, initialized network with novel loss function Lf able guide model better representation features. Then, fine-tuned pretrained networks weighted variant helping improve robustness class imbalance. We evaluated our ISIC 2018 dataset against existing methods imbalance approaches The results demonstrated superiority framework, outperforming all compared by minimum margin 2% single model.