Accelerating Skin Cancer Diagnosis: Harnessing Transfer Learning and CNNs for Multi-Categorical Detection

Skin cancer diagnosis is critical for ensuring effective treatment and improving patient outcomes. To increase detection accuracy over many skin cancer types, this work presents SKINC-Net, an advanced hybrid framework incorporating ensemble learning, transformer-based attention mechanisms, and convolutional neural networks (CNNs). The system extracts multi-level features by means of a hierarchical approach employing a ResNet50 backbone from the ISIC 2020 and ISIC 2021 datasets, therefore capturing both low-level textures and high-level morphological structures. Transformer-based self-attention systems dynamically target diagnostically significant areas, hence enhancing model interpretability and accuracy. The final diagnosis is reached by lowering misclassifications using a multi-categorical ensemble learning method with a confidence-weighted voting system. With high accuracy, experimental findings on the ISIC datasets revealed generally accuracies of 98.4% and 98.1 % respectively, with exceptional precision, recall, and F1-scores. These findings underline the considerable relevance of SKINC-Net as a consistent and interpretable tool for automated skin cancer diagnosis in clinical dermatological settings,