Multi-Modal Framework for Skin Lesion Classification Using Deep Learning and Machine Learning Techniques

Skin cancer, especially melanoma, is one of the most threatening diseases of the modern world. Therefore, accurate and timely diagnosis is critical. This work proposes a methodology based on CNNs and SVM for the dermoscopic and clinical picture classification of skin cancer. The foundation aims to enhance diagnostic consistency using the complementary properties of various imaging techniques. Continuing the tradition of the previous years, HAM10000 and the ISIC 2020 datasets became available to the public. The outcome suggests that the proposed multimodal method CNN-SVM is far better than simple CNN models. The recommended technique achieved a measure of accuracy of 92.5%, precision of91.8%, recall of 92.0%, F1 score of91.9%, and ROC-AUC of 0.95. The CNN model trained on dermoscopic pictures achieved an accuracy of 89.3% and ROC-AUC of 0.92, while the CNN model with clinical images numbered 84.7% and ROC AUC of 0.88%. The improvement of the multimodal method is attributed to the incorporation of complex lesion-specific features from dermoscopic photos and contextual information from clinical images. Thus, the current study focuses on the ability of the SVM to comprehend the features of high-dimensional data and the classification of instances that are often near the decision boundary and closely or equally related to any class. The offered methodology has potential, though small participant numbers and high photo quality limit its applicability. Future work should focus on data extension and management of real-world uncertainty. As such, this study shows that the proposed CNN-SVM multimodal model is a reliable and accurate tool for skin cancer diagnosis in clinical settings.