Multi-institutional PET/CT image segmentation using federated deep transformer learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning models for PET/CT segmentation require large, diverse datasets.
• Sharing multi-institutional data is hindered by legal, ethical, and privacy concerns.
• Federated learning (FL) offers a solution for collaborative model training without direct data sharing.

Purpose of the Study:

• To develop and evaluate a federated learning (FL) framework for multi-institutional PET/CT image segmentation.
• To assess the performance of various FL algorithms against centralized and single-center approaches.
• To address challenges in generalizable and trustworthy deep learning for medical imaging.

Main Methods:

• A dataset of 328 head and neck (HN) cancer patients from six centers was used.
• A transformer network was employed for dual-channel PET/CT image segmentation.
• Seven FL algorithms (ClQu, ZeQu, FedAvg, LoCo, RoAg, SeAg, GDP-AQuCl) were evaluated against centralized and single-center baselines.

Main Results:

• FL algorithms, particularly SeAg and GDP-AQuCl, achieved performance comparable to centralized methods (Dice coefficient 0.80±0.11).
• All FL methods demonstrated relative errors <5% for SUVmax and SUVmean, outperforming single-center baselines.
• No statistically significant performance differences were observed among most FL algorithms.

Conclusions:

• The developed FL framework shows promising performance for HN tumor segmentation in PET/CT images.
• FL enables the creation of generalizable and trustworthy deep learning models despite data privacy constraints.
• Federated learning is a viable approach for multi-institutional medical image analysis.