Prediction of the Microinvasion of Melanoma Using Supervised Deep Convolutional Neural Networks

Melanoma poses a significant global health threat, contributing to over 90% of skin cancer-related deaths. Dermoscopy aids in early melanoma diagnosis, but distinguishing between in situ and invasive melanomas remains challenging, even for experienced dermatologists. Recent strides in artificial intelligence (AI) for medical image analysis suggest its potential to support and offer a secondary opinion to dermatologists in dermoscopy. Utilizing annotated images is crucial for Deep Learning systems to learn associations between images and ground-truth labels. This study employs datasets from diverse sources to train and evaluate Deep Transfer Learning algorithms, specifically using a DenseNet121 convolutional neural network, to the Breslow thickness of melanomas. Supervised learning, employing a stratified 5-fold cross-validation technique, demonstrates strong performance in this melanoma classification task. The research underscores the potential of AI-based automatic diagnosis systems as valuable aids for medical professionals, serving as secondary opinions or triage tools in healthcare settings.