Decentralized Distributed Multi-institutional PET Image Segmentation Using a Federated Deep Learning Framework.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) algorithm generalizability relies on diverse, large datasets.
• Sharing medical images between institutions is hindered by patient privacy, ethical, and legal concerns.
• Multicentric data is crucial for robust medical AI but difficult to aggregate.

Purpose of the Study:

• To develop a federated deep learning (DL) framework for Positron Emission Tomography (PET) image segmentation.
• To evaluate the performance of the federated DL model against a traditional centralized DL approach.
• To address challenges in multicentric medical image data sharing.

Main Methods:

• Utilized PET images from 405 head and neck cancer patients across 9 centers.
• Developed a federated DL model using a modified R2U-Net architecture, comparing it to a centralized model pooling all data.
• Assessed segmentation performance using Dice and Jaccard coefficients and quantitative accuracy of PET parameters (e.g., SUVpeak, SUVmean, MTV, TLG).

Main Results:

• Federated and centralized DL models achieved nearly identical performance for segmentation metrics (Dice: 0.84 ± 0.06 vs 0.84 ± 0.05; Jaccard: 0.73 ± 0.08 vs 0.73 ± 0.07).
• Comparable relative errors (RE%) were observed for quantitative PET parameters like SUVmean, metabolic tumor volume, and total lesion glycolysis between the two methods.
• No statistically significant performance differences were found between the federated and centralized DL frameworks (P > 0.05).

Conclusions:

• The federated DL model demonstrates quantitative performance on par with the centralized DL model for PET image segmentation.
• Federated DL offers a viable solution for creating robust and generalizable segmentation models.
• This approach effectively navigates patient privacy and data sharing restrictions in multicentric clinical research.