Federated Hybrid CNN Vision Transformer Framework for Breast Cancer Classification Under Simulated Non-IID Settings
ASIDE Oncology · 09.03.2026 · 1 (1) · 28-37
DOI: 10.71079/aside.onc.030926535 ↗Kurz gefasst
Fachartikel in der Zeitschrift ASIDE Oncology, online erschienen am 09.03.2026. Verfasst von Bandhan Panda und 3 weiteren Autor:innen. Die Arbeit nennt 19 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Background
Breast cancer diagnosis increasingly relies on data-driven learning from heterogeneous medical imaging sources. However, centralized deep learning approaches face major limitations, including privacy risks, institutional data silos, and limited generalization across imaging modalities.
Methods
This study proposes a privacy-aware federated learning framework integrating a hybrid CNN–Vision Transformer architecture for breast cancer classification under simulated non-identically distributed (non-IID) conditions. Public datasets representing different imaging modalities, BreakHis (histopathology), INbreast and CBIS-DDSM (mammography), and BUSI (ultrasound) are treated as independent federated clients to emulate multi-institutional collaboration. Each client trains a local model, and only the model parameters are aggregated via Federated Averaging, thereby preserving data locality. The hybrid architecture combines a ResNet-18 convolutional branch for local feature extraction with a Vision Transformer branch for global contextual representation.
Results
Across ten federated communication rounds, the global model demonstrates stable convergence under heterogeneous client distributions. The final global validation performance reaches 72.43% accuracy, AUC 0.7475, and F1-score 0.718. Evaluation on the pooled test cohort achieves approximately 84% overall accuracy with a weighted F1-score of 0.82, while malignant recall approaches 96%, prioritizing clinically critical cancer detection. Qualitative explainability analysis indicates that the model focuses on diagnostically relevant tissue regions.
Conclusion
The results demonstrate the feasibility of hybrid CNN–Vision Transformer training in a federated setting for privacy-aware breast cancer classification across heterogeneous imaging domains.
Zitieren
APA
Panda, B., Patro, B. K., Das, S. S., & Kar, S. K. (2026). Federated Hybrid CNN Vision Transformer Framework for Breast Cancer Classification Under Simulated Non-IID Settings. ASIDE Oncology, 1(1), 28-37. https://doi.org/10.71079/aside.onc.030926535
Vancouver
Panda B, Patro BK, Das SS, Kar SK. Federated Hybrid CNN Vision Transformer Framework for Breast Cancer Classification Under Simulated Non-IID Settings. ASIDE Oncology. 2026;1(1):28-37. doi:10.71079/aside.onc.030926535
BibTeX
@article{Panda2026,
author = {Panda, Bandhan and Patro, Bibek Kumar and Das, Siba Sundar and Kar, Santosh Kumar},
title = {Federated Hybrid CNN Vision Transformer Framework for Breast Cancer Classification Under Simulated Non-IID Settings},
journal = {ASIDE Oncology},
year = {2026},
volume = {1},
number = {1},
pages = {28--37},
doi = {10.71079/aside.onc.030926535}
}
EUVIMED