Abstract
Federated Learning (FL) has emerged as a promising solution to address the limitations of centralised machine learning (ML) in oncology, particularly in overcoming privacy concerns and harnessing the power of diverse, multi-center data. This systematic review synthesises current knowledge on the state-of-the-art FL in oncology, focusing on breast, lung, and prostate cancer. Distinct from previous surveys, our comprehensive review critically evaluates the real-world implementation and impact of FL on cancer care, demonstrating its effectiveness in enhancing ML generalisability, performance and data privacy in clinical settings and data. We evaluated state-of-the-art advances in FL, demonstrating its growing adoption amid tightening data privacy regulations. FL outperformed centralised ML in 15 out of the 25 studies reviewed, spanning diverse ML models and clinical applications, and facilitating integration of multi-modal information for precision medicine. Despite the current challenges identified in reproducibility, standardisation and methodology across studies, the demonstrable benefits of FL in harnessing real-world data and addressing clinical needs highlight its significant potential for advancing cancer research.
Introduction: Breaking the Data Silos
In oncology, data is the most valuable asset for developing precision medicine, yet it is often siloed within high-security institutional vaults. Training robust Machine Learning (ML) models requires massive, diverse datasets to ensure they generalize across different demographics and clinical settings. Traditionally, this required "centralized learning"—gathering all data into a single repository. However, this approach frequently encounters barriers such as complex legal agreements and strict privacy regulations like GDPR or HIPAA.
Federated Learning (FL) offers a decentralized alternative: instead of moving data to the model, the model travels to the data. This systematic review evaluates the real-world impact of FL across breast, lung, and prostate cancer research, investigating whether this approach can effectively bridge the gap between data privacy and model performance.
Methodology and Technical Landscape
Following PRISMA guidelines, the authors identified 25 key papers published between 2020 and 2023. To evaluate the quality of these studies, they introduced an "FL rigour" assessment based on two criteria:
- Comparative Framework: Whether the study compared the FL model against a centralized ML baseline using the same data.
- Performance: Whether the FL model surpassed or matched the performance of those baselines.
The review analyzed a wide spectrum of architectures, finding that Large Pre-trained models and CNNs are the most frequently utilized tools in the current federated landscape.
Clinical Applications and the "Privacy Bonus"
While FL is often adopted primarily as a privacy-preserving workaround, the review highlights a "privacy bonus": FL frequently leads to better models. By enabling access to diverse, multi-center data that would otherwise be inaccessible, FL models learn more representative patterns and reduce overfitting.
As shown in the figures above, tumor identification and classification remain the primary focus of current research. However, FL's unique ability to integrate disparate data types—such as combining Whole-Slide Imaging (WSI) pathology with clinical Electronic Health Record (EHR) data—positions it as a cornerstone for future multi-modal precision medicine.
Key Results: Outperforming the Center
The most significant finding of the review is that FL is not just a compromise. In 60% of the studies (15/25), FL actually outperformed centralized ML. Key performance benchmarks include:
- Breast Cancer: 95.95% accuracy using federated averaging on histopathology images.
- Lung Cancer: 99.69% accuracy in detection, outperforming centralized methods while reducing training time.
- Prostate Cancer: A Dice score of 89.5% for CT segmentation on private data.
FL models consistently demonstrated higher performance on "unseen" data from different institutions compared to models trained locally at a single site, proving their superior generalizability.
Challenges and Future Directions
Despite its potential, the field faces several "growing pains":
- Reproducibility: Only 8 of the 18 "rigorous" papers provided accessible code, hindering clinical adoption.
- Standardization: The lack of universally accepted benchmark datasets makes it difficult to compare different federated algorithms fairly.
- Communication Overhead: The "chatter" of sending model weights back and forth can be significant, especially with large foundation models.
Future research should focus on advanced aggregation methods (like FedProx or FedNova) to handle heterogeneous data across hospitals and explore the integration of Foundation Models within federated frameworks to fully realize the transformative power of FL in cancer care.
Summary updated