Evolución de Arquitecturas Híbridas CNN-Transformer y Técnicas de Explicabilidad aplicadas al diagnóstico dermatológico mediante Deep Learning

Skin cancer is the most frequently diagnosed malignant neoplasm worldwide, with mortality rates that increase dramatically when melanoma is detected at advanced stages. Early diagnosis assisted by artificial intelligence systems represents a critical strategy to reduce such mortality. This scientific research reviews and synthesizes the state of the art in hybrid CNN-Transformer architectures applied to dermatological diagnosis using Deep Learning, with special emphasis on explainability techniques (XAI). Thirty scientific articles were analyzed through the PRISMA protocol from high-impact academic databases (PubMed, IEEE Xplore, Scopus, Web of Science). Results reveal an evolution across three generations: classical CNNs (83–88% balanced accuracy), pure Vision Transformers (88–91%), and hybrid CNN-Transformer architectures (90–94%), the latter being the most active research frontier. A methodology is proposed for an EfficientNet-B4 + Swin Transformer ensemble model with integrated XAI module (Grad-CAM++ + SHAP) on the HAM10000 dataset, with an expected balanced accuracy exceeding 92% with high clinical interpretability.