Early detection and accurate identification of melanoma, the most lethal form of skin cancer, are critical for improving patient survival rates. However, conventional diagnostic tools often fail to detect early-stage melanoma due to limited sensitivity. This research introduces an advanced Vision Transformer (ViT)-based model equipped with explainability features, specifically designed to address the complex challenge of early-stage melanoma detection. Self-attention-based Melanoma Analysis using Reliable Transformers with Explainable Artificial Intelligence (SMART-XAI) combines ViT with self-attention methods and explainable AI techniques like Attention Rollout and Self-Attention Attribution to create a melanoma diagnostic system. The ViT architecture leverages multi-scale feature extraction and self-attention mechanisms to capture both local and global patterns within skin lesion images. By segmenting images into patches, the model effectively identifies critical melanoma features across various scales, enhancing classification accuracy and interpretability. To ensure clinical transparency, the model incorporates explainable artificial intelligence (AI) techniques, namely Attention Rollout and Self-Attention Attribution, which enable the visualization of image regions that influence the modelís decisions. This interpretability allows clinicians to understand the diagnostic rationale behind AI, thereby fostering greater trust and usability in clinical settings. Experimental validation demonstrated strong performance, achieving 96.0 ± 0.87% accuracy, 95.2 ± 0.75% sensitivity, and 94.1 ± 1.05% specificity across 5-fold cross-validation. These results confirm the system’s reliability as an early detection tool for melanoma, while providing interpretable insights for informed clinical decision-making. Overall, the integration of deep learning with explainable AI represents a significant advancement in dermatological diagnostics, particularly for the management of early-stage melanoma.
Mehdar et al. (Fri,) studied this question.