Related Experiment Videos
Federated learning for privacy-preserving skin cancer classification using deep neural networks
Mohammed A M Alfalahi1,2, Oğuz Karan3, Sefer Kurnaz1
1Electrical and Computer Engineering Department, Engineering College, Altinbas University, Istanbul, Türkiye.
Frontiers in Oncology
|May 13, 2026
Summary
Federated learning (FL) enables effective skin cancer classification without compromising patient privacy. A decentralized ring-based approach using MobileNetV2 achieved 98.88% accuracy, outperforming centralized methods.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computational Biology
Background:
- Skin cancer diagnosis requires precise, scalable, and privacy-preserving systems.
- Centralized deep learning is limited by data-sharing restrictions and institutional heterogeneity.
- Federated learning (FL) offers a solution by enabling collaborative model training without raw data transfer.
Purpose of the Study:
- To implement and evaluate a privacy-sensitive federated learning architecture for skin cancer classification in decentralized clinical settings.
- To compare centralized Federated Averaging (FedAvg) with a decentralized ring-based peer-to-peer strategy.
- To assess the framework's robustness against non-IID data distributions and institutional heterogeneity.
Main Methods:
- Developed a heterogeneity-sensitive client partitioning algorithm using Kullback-Leibler divergence.
- Compared MobileNetV2 and VGG16 backbone architectures with centralized FedAvg and a decentralized ring topology.
- Evaluated performance against FedProx and SCAFFOLD across various non-IID severity levels on ISIC and Skin Cancer: Malignant vs. Benign datasets.
- Incorporated differential privacy to balance model utility and privacy.
Main Results:
- The decentralized ring-based FL with MobileNetV2 achieved the highest performance, with 98.88% accuracy and 98.80% F1-score.
- This configuration significantly outperformed the centralized baseline (p < 0.01) and demonstrated superior resistance to heterogeneous data.
- MobileNetV2 showed faster convergence, reduced communication overhead, and better performance compared to VGG16 in federated settings.
Conclusions:
- Federated learning is as effective, if not more so, than centralized learning for skin cancer classification while preserving patient privacy.
- Lightweight architectures like MobileNetV2 are well-suited for FL due to efficient gradient propagation and low communication costs.
- The proposed framework offers a feasible, regulation-compliant solution for AI-aided dermatological diagnosis in distributed healthcare environments.
