Federated Learning with Research Prototypes: Application to Multi-Center MRI-based Detection of Prostate Cancer with...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Prostate cancer diagnostics

Background:

• Prostate cancer detection and staging using MRI presents significant challenges for both radiologists and deep learning algorithms.
• Improving the performance of these algorithms requires access to large, diverse datasets, which are often siloed within individual institutions.

Purpose of the Study:

• To introduce a flexible federated learning framework for the cross-site training, validation, and evaluation of deep learning models for prostate cancer detection.
• To enable prototype-stage algorithms to leverage multi-institutional data while preserving patient privacy and data security.

Main Methods:

• Developed a novel abstraction of prostate cancer groundtruth to accommodate diverse annotation and histopathology data.
• Utilized UCNet, a custom 3D UNet, for simultaneous supervision of pixel-wise, region-wise, and gland-wise classification.
• Implemented cross-site federated training on over 1400 heterogeneous multi-parametric MRI exams from two university hospitals.

Main Results:

• Achieved significant improvements in cross-site generalization performance for both lesion segmentation and per-lesion binary classification of clinically significant prostate cancer.
• Observed a 100% improvement in cross-site lesion segmentation intersection-over-union (IoU).
• Reported an overall accuracy improvement of 9.5-14.8% in cross-site lesion classification.

Conclusions:

• Federated learning effectively enhances the generalization performance of prostate cancer detection models across institutions.
• The framework protects patient health information and institution-specific data.
• An open-source FLtools system is provided to facilitate the adoption of federated learning in medical imaging projects.