AI Skin Cancer Detector Hits 98% Accuracy with Multispectral Metasurface

Skin cancer remains one of the most common malignancies globally, and early detection is critical. Conventional dermoscopic techniques are limited to the visible spectrum, often missing subtle spectral signatures of early-stage malignancies. To address this, Afsane Saee Arezoomand and team propose an intelligent detection framework that combines a multispectral metasurface with a hybrid deep learning model. The metasurface acts as a non-invasive imaging tool, capturing spectral data across multiple wavelengths sensitive to tissue alterations. This data is then processed by a hybrid architecture of Convolutional Neural Networks (CNNs) for local features and Vision Transformers (ViTs) for global context, enabling robust classification of skin lesions.

In simulation-based evaluations, the framework achieved approximately 98% accuracy, 95% sensitivity, and 99% specificity — outperforming conventional RGB-based and single-architecture approaches. Attention maps revealed the model focuses on clinically relevant lesion regions, enhancing interpretability for clinicians. The authors suggest this combination could lead to portable, fast, and highly accurate diagnostic tools for dermatology. Published in New Researches in the Smart City and on arXiv, the work highlights a promising path for AI-powered medical imaging, potentially enabling earlier and more reliable skin cancer detection.