Multimodal deep learning ensemble framework for skin cancer detection

Abstract

Skin cancer is the abnormal growth of skin cells, most often developing on skin exposed to the sun. It is among the most fatal forms of cancer, making its early detection and therapy crucial. In addition to conventional techniques, deep learning methods are increasingly utilized for accurate identification and classification. This study proposes a convolutional neural network (CNN) model using transfer learning to detect and classify multiple types of skin cancer. This study focuses on evaluating the efficacy of transfer learning techniques in enhancing CNN performance for this critical task. A significant contribution of this study is the use of transfer learning to improve CNN performance in skin cancer detection and classification by leveraging pre-trained models, including ResNet50, Xception, MobileNet, EfficientNetB0, and DenseNet121. The integration of metadata demonstrated a significant improvement in accuracy compared to using images alone, enhancing the performance of most models. Further enhancement was achieved through ensemble techniques, specifically an adaptive weighted ensemble method, which dynamically assigns weights to individual models based on their performance, resulting in superior overall accuracy. SMOTE was used as an oversampling technique to address class imbalance. The proposed fusion of pre-trained models (ResNet50, Xception, and EfficientNetB0) combined with metadata achieved 93.2% accuracy, 93% precision, 93% recall, 93% F1 score, and 97.3% AUC on the ISIC 2018 dataset. On the ISIC 2019 dataset, it achieved 91.1% accuracy, 92% precision, 93% recall, 92% F1 score, and 95.5% AUC, surpassing many state-of-the-art methods. Experiments on an external dataset, Derm7pt, resulted in 82.5% accuracy, with a precision of 86%, recall of 83%, F1 score of 84% and AUC of 89.15%, demonstrating the improved interpretability and generalization of the proposed model. The proposed ensemble model optimizes deep learning for healthcare applications, enhancing dermatological diagnosis and treatment strategies for skin cancer patients.