EUVIMEDEUVIMED entdecken
Crossref

Explainable Deep Learning for Dermoscopic Skin Lesion Classification: A Comparative Study of Vision Transformers and Convolutional Networks with Grad-CAM

Sidra Hameed, Muhammad Nauman, Muhammad Hasnain, Muhammed Mamoon, Zeshan Afzal, Kainat Bukhari

VFAST Transactions on Software Engineering · 25.06.2026 · 14 (2) · 318-342

DOI: 10.21015/vtse.v14i2.2419 ↗

Kurz gefasst

Fachartikel in der Zeitschrift VFAST Transactions on Software Engineering, online erschienen am 25.06.2026. Verfasst von Sidra Hameed und 5 weiteren Autor:innen. Die Arbeit nennt 71 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Skin cancer is a major global health burden, and computer-aided systems are increasingly attractive for improving screening and diagnosis through accurate and consistent analysis of dermoscopic images. Deep learning models can achieve high performance in skin lesion classification, but their black-box nature limits clinicians’ ability to understand individual predictions in high-stakes clinical contexts. Explainable Artificial Intelligence (XAI) offers a way to expose model reasoning in human-interpretable form, yet further enhancements are needed to support reliable and trustworthy AI-assisted diagnosis. This study proposes an XAI-driven framework for dermoscopic skin lesion classification using state-of-the-art deep learning models combined with Gradient-weighted Class Activation Mapping (Grad-CAM) to improve interpretability. Dermoscopic images from a public dataset are preprocessed through resizing, normalization, and augmentation, and several architectures are trained and evaluated, including a Vision Transformer (ViT), a baseline convolutional neural network (CNN), ResNet-18, ResNet-50, and DenseNet-121. Grad-CAM is applied to each model to generate heatmaps that highlight the image regions most influential for the predicted classes. The results show that transformer-based and ResNet-based models achieve competitive accuracy and strong macro-averaged performance across multiple lesion categories. Grad-CAM visualizations indicate that convolutional models tend to focus on localized lesion regions, whereas the ViT model captures more global contextual patterns. Overall, combining transformer and convolutional architectures with Grad-CAM yields improved interpretability without compromising predictive performance, supporting more credible clinical decision support and providing a transferable framework for future XAI research in dermatology.

Zitieren

APA

Hameed, S., Nauman, M., Hasnain, M., Mamoon, M., Afzal, Z., & Bukhari, K. (2026). Explainable Deep Learning for Dermoscopic Skin Lesion Classification: A Comparative Study of Vision Transformers and Convolutional Networks with Grad-CAM. VFAST Transactions on Software Engineering, 14(2), 318-342. https://doi.org/10.21015/vtse.v14i2.2419

Vancouver

Hameed S, Nauman M, Hasnain M, Mamoon M, Afzal Z, Bukhari K. Explainable Deep Learning for Dermoscopic Skin Lesion Classification: A Comparative Study of Vision Transformers and Convolutional Networks with Grad-CAM. VFAST Transactions on Software Engineering. 2026;14(2):318-342. doi:10.21015/vtse.v14i2.2419

BibTeX
@article{Hameed2026,
  author = {Hameed, Sidra and Nauman, Muhammad and Hasnain, Muhammad and Mamoon, Muhammed and Afzal, Zeshan and Bukhari, Kainat},
  title = {Explainable Deep Learning for Dermoscopic Skin Lesion Classification: A Comparative Study of Vision Transformers and Convolutional Networks with Grad-CAM},
  journal = {VFAST Transactions on Software Engineering},
  year = {2026},
  volume = {14},
  number = {2},
  pages = {318--342},
  doi = {10.21015/vtse.v14i2.2419}
}
Kostenlos starten →