MiSC: A hybrid multi-modal deep learning approach for accurate skin cancer detection
Ensaf Hussein Mohamed, Nada Abdu, Mostafa M.H. Khalil, Hossam Kamal, Essam A. Rashed
Abstract
Abstract With skin cancer rates increasing globally due to factors like prolonged ultraviolet exposure, early and accurate detection methods are crucial in managing and mitigating the disease. Implementing innovative diagnostic techniques can significantly enhance early intervention, improving patient outcomes and survival rates. This research aims to precisely identify and classify six skin cancer types. The proposed method integrates clinical images and metadata into a hybrid deep learning model, leveraging feature extraction and classification. The proposed approach employs seven machine learning algorithms alongside ten pre-trained deep learning models, notably featuring MobileNetV2 for image feature extraction, logistic regression for image classification, and random forest for metadata analysis. The developed approach is tested with the PAD-UFES-20 dataset and it demonstrate a notable improvement in diagnostic accuracy (95.6%), precision (96.8%), recall (95.6%), and F1-Score (95.7%). This findings highlight the significant contribution of metadata in enhancing classification accuracy, surpassing image-based methods alone.