Deep Dermatology: Enhancing Skin Cancer Precision through Adaptive Transfer Learning

https://doi.org/10.5281/zenodo.19467937

Abstract

Skin cancer classification requires accurate and efficient models for early diagnosis to reduce mortality. While deep learning techniques show promise, achieving a balance between diagnostic speed and precision remains a challenge. This study evaluates three deep learning models—Convolutional Neural Network (CNN), EfficientNetV2-B0, and Vision Transformer (ViT-B16)—to identify optimal architectures for clinical deployment. Using Kaggle’s ISIC Skin Cancer Detection Dataset (2,637 training images and 660 test images), we applied transfer learning, data augmentation, and a suite of performance metrics including accuracy, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC). The results demonstrated that transfer learning models significantly outperformed the baseline CNN. EfficientNetV2 achieved 86.67% accuracy with an MCC of 0.7307 and rapid inference time (0.00190 seconds per sample), making it suitable for real-time diagnostics. Vision Transformer (ViT-B16) attained the highest accuracy of 88.48% and MCC of 0.7675 but had slower inference (0.00852 seconds per sample). These findings indicate that EfficientNetV2 provides the best balance between accuracy and speed for resource-constrained clinical settings, while ViT-B16 prioritizes diagnostic precision. This research provides valuable insights for context-specific model selection in dermatology, highlighting the potential of transfer learning to improve skin cancer detection. Future work will explore hybrid architectures to optimize the accuracy-speed trade-off.