Multi-Scale Ensemble Deep Learning for Automated Skin Lesion Classification Using ResNet-18: A Dermatoscopic Study on Ham10000
International Journal of Drug Delivery Technology · 01.06.2026 · 16 (42s)
DOI: 10.25258/ijddt.16.42s.9 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 01.06.2026. Verfasst von K. Vinotha und 6 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Skin cancer remains one of the most prevalent malignancies globally, with early and accurate detection being critical to patient survival. Dermoscopic imaging combined with deep learning offers a promising non-invasive diagnostic pathway.
Objective
This study proposes a Multi-Scale Ensemble Inference framework built upon a transfer-learned ResNet-18 backbone to classify seven categories of skin lesions from the HAM10000 dataset with high accuracy and robustness against class imbalance.
Methods
The proposed system employs a dual-stream inference pipeline: a global stream processing full 224x224 dermoscopic images and a localized zoom stream that centre-crops to 176x176 pixels before upsampling, mimicking a dermatologist's magnification workflow. Weighted Cross-Entropy Loss (WCE), AdamW optimisation, temperature scaling (T=1.1), and Test-Time Augmentation (TTA) are integrated into a unified training pipeline.
Results
The proposed framework achieves a validated accuracy of 89.57% on the HAM10000 test set, outperforming ResNet-50 (78.40%), DenseNet-121 (84.50%), and prior ensemble approaches (85.80%) reported in the literature. The Multi-Scale Ensemble stage alone contributed a +4.79% accuracy gain over the AdamW-optimised baseline.
Conclusion
The study demonstrates that lightweight architectures supplemented with multi-scale inference and probabilistic calibration can achieve state-of-the-art performance on imbalanced medical imaging datasets, making the approach suitable for real-time clinical deployment.
Zitieren
APA
Vinotha, K., Vigneshkumar, K., Balamurugan, T., Ajay, C., Sudarsanam, M., Dhinesh, K., & Jibin, E. (2026). Multi-Scale Ensemble Deep Learning for Automated Skin Lesion Classification Using ResNet-18: A Dermatoscopic Study on Ham10000. International Journal of Drug Delivery Technology, 16(42s). https://doi.org/10.25258/ijddt.16.42s.9
Vancouver
Vinotha K, Vigneshkumar K, Balamurugan T, Ajay C, Sudarsanam M, Dhinesh K, et al. Multi-Scale Ensemble Deep Learning for Automated Skin Lesion Classification Using ResNet-18: A Dermatoscopic Study on Ham10000. International Journal of Drug Delivery Technology. 2026;16(42s). doi:10.25258/ijddt.16.42s.9
BibTeX
@article{Vinotha2026,
author = {Vinotha, K. and Vigneshkumar, K. and Balamurugan, T. and Ajay, C. and Sudarsanam, M. and Dhinesh, K. and Jibin, E.},
title = {Multi-Scale Ensemble Deep Learning for Automated Skin Lesion Classification Using ResNet-18: A Dermatoscopic Study on Ham10000},
journal = {International Journal of Drug Delivery Technology},
year = {2026},
volume = {16},
number = {42s},
doi = {10.25258/ijddt.16.42s.9}
}
EUVIMED