DeepSkin: A Deep Learning Approach to Skin Cancer Classification

Abstract

This study addresses the urgent global problem of rapidly spreading skin cancer, emphasizing the significance of accurate detection for effective prevention. Dermatologists are utilizing deep learning, namely Convolutional Neural Networks (CNNs), because of their difficulties with early detection. The work employs data preparation techniques such autoencoder-based segmentation, dull razor, and sampling using the MNIST: HAM10000 dataset, which has 10,015 samples and seven types of skin diseases. It is demonstrated that when DenseNet169 and ResNet50 models are employed for transfer learning, DenseNet169's undersampling results in high accuracy and F1- measure, whereas ResNet50's oversampling technique works well in both measures. Building on the primary paper's use of ResNet50, DenseNet161, and VGG16 (achieving 91% accuracy), this extension explores additional models including Xception, DenseNet201, and InceptionV3. The study shows how different models and parameter modifications may improve the classification of skin cancer, offering a workable way to increase diagnosis accuracy and preventative measures. It expects an increase in accuracy of 95%.