Federated learning enables big data for rare cancer boundary detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Machine learning (ML) generalizability is limited by data silos.
• Centralizing multi-site data for ML is often infeasible.
• Federated ML (FL) offers a privacy-preserving alternative by sharing model updates.

Purpose of the Study:

• To conduct the largest Federated ML (FL) study for glioblastoma tumor boundary detection.
• To assess the efficacy of FL in improving ML model generalizability across diverse, multi-site datasets.
• To establish a new paradigm for large-scale, multi-institutional healthcare research.

Main Methods:

• Utilized FL with data from 71 international sites, encompassing 6,314 glioblastoma cases.
• Developed an automatic tumor boundary detection model using FL.
• Compared FL model performance against a publicly trained model.

Main Results:

• Achieved a 33% improvement in delineating the surgically targetable tumor extent.
• Demonstrated a 23% improvement in delineating the complete tumor extent.
• Successfully scaled FL to a large, multi-continental dataset for a rare disease.

Conclusions:

• FL is effective for developing accurate and generalizable ML models in healthcare, even for rare diseases.
• This study demonstrates FL's potential to overcome data-sharing barriers and enable research on diverse populations.
• The findings suggest FL as a paradigm shift for future multi-site medical research and collaboration.