DeepSkin: A Deep Learning Method for Categorizing Skin Cancer

Abstract

This study tackles the pressing worldwide issue of fast spreading skin cancer, highlighting how important a precise diagnosis is to successful prevention. Due of their challenges with early detection, dermatologists are using deep learning, namely Convolutional Neural Networks (CNNs). Utilizing the MNIST: HAM10000 dataset, which has 10,015 samples and seven kinds of skin lesions, the study uses data preparation methods such as autoencoderbased segmentation, dull razor, and sampling. When DenseNet169 and ResNet50 models are used for transfer learning, it is shown that while ResNet50's oversampling strategy performs well in both metrics, DenseNet169's undersampling produces high accuracy and F1-measure. This expansion investigates other models such as Xception, DenseNet201, and InceptionV3, building on the main paper's usage of ResNet50, DenseNet161, and VGG16 (reaching 91% accuracy). The study highlights the potential of various models and parameter adjustment to enhance skin cancer categorization, providing a viable path for improving diagnostic precision and preventive tactics. It anticipates a 95% accuracy increase