Machine Learning for Early Detection of Melanoma Using Clinical and Image Data

Abstract

Melanoma is one of the most aggressive skin cancers, and early detection remains a key factor in improving survival outcomes. However, all current methods are plagued by heterogeneous data formats, class imbalance, and limited interpretability. In this respect, the paper proposes MH-CNNXNet, a multimodal, hybrid model that merges dermoscopic image features with structured clinical metadata for more accurate melanoma detection. Our work leverages the massive dataset of 33,126 images from the SIIM-ISIC dataset and corresponding clinical variables to provide a model that can effectively fuse Convolutional Neural Network (CNN)-based visual embeddings with an Extreme Gradient Boosting (XGBoost)-derived clinical representation supported by extensive preprocessing, augmentation, and class-balanced optimization. To this end, experimental results show that MH-CNNXNet outperforms all baselines of clinical and image-only models, including ResNet18 and XGBoost, with 99.56% accuracy, 99.56% precision, and 99.53% recall. Clinical trust is further enhanced by a SHapley Additive explanation (SHAP)-based explainability layer underlining clinical and visual cues of biological significance. This study presents notable improvements in diagnostic performance and reliability over earlier single-modality approaches and creates a generally scalable, interpretable, clinically aligned framework for early melanoma detection.