Using federated data sources and Varian Learning Portal framework to train a neural network model for automatic organ...

Area of Science:

• Medical Imaging
• Machine Learning
• Radiotherapy

Background:

• Federated learning enables multi-institutional research without direct data sharing.
• Privacy-preserving machine learning is crucial for sensitive medical data.
• Accurate organ segmentation is vital for radiotherapy planning.

Purpose of the Study:

• To train a deep neural network for female pelvic organ segmentation using a distributed framework.
• To evaluate the prediction power of a federated learning model against a centralized model.
• To assess the feasibility of the Varian Learning Portal (VLP) for privacy-preserving medical research.

Main Methods:

• Utilized the Varian Learning Portal (VLP), a distributed machine learning infrastructure.
• Trained a deep neural network model for female pelvic organ segmentation across multiple hospital sites.
• Employed a synchronous data distributed framework for privacy-preserving model training.

Main Results:

• The federated learning model achieved prediction performance on par with a centralized training model.
• The VLP infrastructure facilitated effective GPU-based training for complex organ segmentation.
• The study demonstrated successful organ segmentation for the female pelvic region.

Conclusions:

• The Varian Learning Portal (VLP) supports privacy-preserving, GPU-based deep learning for organ segmentation.
• Federated learning via VLP allows training on diverse datasets from multiple clinics without compromising patient privacy.
• This approach is effective for challenging segmentation tasks like the female pelvic region, improving model robustness.