Federated Learning Enables Big Data for Rare Cancer Boundary Detection (2204.10836v2)

Abstract: Although ML has shown promise in numerous domains, there are concerns about generalizability to out-of-sample data. This is currently addressed by centrally sharing ample, and importantly diverse, data from multiple sites. However, such centralization is challenging to scale (or even not feasible) due to various limitations. Federated ML (FL) provides an alternative to train accurate and generalizable ML models, by only sharing numerical model updates. Here we present findings from the largest FL study to-date, involving data from 71 healthcare institutions across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, utilizing the largest dataset of such patients ever used in the literature (25,256 MRI scans from 6,314 patients). We demonstrate a 33% improvement over a publicly trained model to delineate the surgically targetable tumor, and 23% improvement over the tumor's entire extent. We anticipate our study to: 1) enable more studies in healthcare informed by large and diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further quantitative analyses for glioblastoma via performance optimization of our consensus model for eventual public release, and 3) demonstrate the effectiveness of FL at such scale and task complexity as a paradigm shift for multi-site collaborations, alleviating the need for data sharing.

• The paper demonstrates a large-scale federated learning application involving 71 institutions across six continents to train models for rare cancer boundary detection, using glioblastoma as a case study.
• This federated approach significantly improved tumor boundary delineation by 33% and 23% compared to models trained on a smaller, centrally hosted dataset.
• Federated learning facilitates extensive collaborative medical research without requiring sensitive data centralization, addressing privacy concerns and enabling the use of diverse, large datasets.

Overview of the Paper: Federated Learning Enables Big Data for Rare Cancer Boundary Detection

The paper explores the application of Federated Learning (FL) in the context of medical imaging, specifically focusing on rare cancer boundary detection, with glioblastoma serving as the primary use case. This paper represents the largest real-world FL effort in healthcare, encompassing data collected from 71 healthcare institutions across six continents to train machine learning models for automatic tumor boundary detection. The results indicate significant improvements in delineating tumor boundaries over centrally trained models, demonstrating federated learning's capacity to bridge the gap in data accessibility while maintaining high privacy standards.

Key Findings

• Improvement in Model Performance: The federated approach led to a 33% and 23% improvement in delineating surgically targetable tumors and the entire tumor extent, respectively, when compared to models trained on a smaller, centrally-hosted dataset.
• Scale and Diversity: The paper utilizes data from 25,256 MRI scans of 6,314 patients, forming the largest dataset for glioblastoma boundary detection reported in the literature. This extensive dataset allows for more diverse and representative training, enhancing model generalizability.
• Data Privacy and Collaboration: FL facilitates collaboration across geographically distinct sites without requiring data centralization, addressing privacy, data ownership, and compliance issues prevalent in multi-site collaborations.
• Potential Clinical Impact: By enabling more extensive data-driven studies without centralized access to sensitive information, the paper paves the way for improved quantitative analyses in glioblastoma, promising to enhance patient-specific treatment planning.

Implications and Future Developments

The paper underscores the utility of FL in supporting large-scale collaborative research in medical applications, particularly for conditions with limited data availability, such as glioblastoma. By mitigating the need for data sharing, FL can potentially foster more collaborative healthcare studies, incentivize data diversity, and enhance data quality, which are crucial for developing accurate and generalizable models. As healthcare systems worldwide continue to adopt AI-driven solutions, FL provides a viable pathway to realize the full benefits of big data analytics in clinical settings while maintaining data integrity and confidentiality.

Future research could focus on refining FL processes to further optimize model performance in low-resource clinical environments and expanding its application to other rare medical conditions. The methodological advancements demonstrated here offer a foundation upon which broader investigations into federated learning applications for complex and privacy-sensitive datasets can be constructed.