Publications

Abstract. Skin cancer detection using deep learning has shown promise, but clinical adoption is what lacks addressing. This study addresses three critical gaps which give confidence for medical adoption. Lack of uncertainty quantification, poor minority class performance and high computational barriers. We developed a swin transformer system using HAM10000 datasets (10, 015 dermascopic images, 7 classes). Our approach integrates Monte carlo dropout for uncertainty estimation, triple -strategy imbalance handling weighted sampling, class-weighted focal loss, and memory-efficient optimization techniques. The model achieved 87.82% test accuracy with 90.15% validation accuracy. Through selective prediction, accuracy on high-confidence cases (80% coverage) reached 97% while uncertain cases (20%) were flagged for expert review. Minority class F1-scores averaged 83.8% with the rarest class achieving 95.7% Memory optimizations reduced peak VRAM to 8GB, enabling training on consumer hardware. These results demonstrate that swin transformers can be production-ready for dermatology when combined with uncertainty quantification and resource-efficient training strategies, providing a practical framework for clinical AI deployment. Keywords- skin cancer detection, Swin Transformer, uncertainty quantification, Monte Carlo Dropout.