Deep DenseNet Architectures for Multi-Class Skin Cancer Classification Using Dermoscopic Images

Abstract

Skin cancer (SC) is one of the most common and rapidly increasing forms of cancer around the world. Early and accurate diagnosis is productive if want to improve treatment outcomes and reduce mortality. Deep learning (DL) techniques have shown great potential for medical image analysis in the past few years, including automatic skin lesion classification. Deep DenseNet Architectures for Multi-Class Skin Cancer Classification with Dermoscopic Images: An Exploratory Study Abstract Melanoma is one of the most dangerous form of skin cancer that originates from melanocytes in the skin layer. Transfer learning using weights from ImageNet trained three DenseNet models (DenseNet121, DenseNet169, and DenseNet201). Train and test datasets: The ISIC dermoscopic skin lesion dataset wsa used to train and run tests of the models. In order to enhance their generalisation, dropout regularisation and learning rate scheduling were used during training. The models was evaluated utilizing a number of performance measures, counting accuracy, precision, recall and F1-score. The results shows the DenseNet201 achieved the best classification performance with an accuracy of 90.69%, a precision of 90.75%, a recall of 90.49% and an F1-score of 90.52%. The ROC and Precision–Recall curves also show the model's strong ability to discriminate among the various SC classes. A comparative analysis with some of the most recent state-of-the-art methods shows that the proposed DenseNet-based approach offers competitive performance while keeping the architecture simple. The findings point to the potential of deep convolutional neural networks (CNN) as effective tools for helping dermatologists with automated SC diagnosis.