DEEP LEARNING APPROACHES FOR EARLY DETECTION OF SKIN CANCER

Melanoma is one of the most aggressive skin cancer types and is notorious for rapid progression, highlighting the necessity of early detection to increase survival outcomes. The conventional diagnostic process can be invasive, consumes too much time, and is relatively dependent on professional evaluation. In this study, the following deep learning models are evaluated: VGG19, DenseNet121, InceptionV3, and EfficientNetB1 with various datasets such as HAM10000, PAD-UFES-20, and ISIC 2016. The deep learning models are evaluated as a function of accuracy and ROC-AUC metrics, with consideration of other problems such as class imbalance, computational performance, and model complexity. The study discusses processes for data augmentation, efficient image processing, as well as ways to incorporate clinical aspects, which could increase the accuracy of skin cancer diagnosis by deep learning models. Future work consists of investigating efficient lightweight models and conducting clinical evaluations to allow real-world applications.