Multi-site distributed training with data protections for PET-based synthetic CT.

Abstract

BACKGROUND: Accurate PET quantification relies on attenuation correction (AC), commonly performed using a linear attenuation map derived from a CT acquisition. However, CT can introduce misregistration artifacts and adds radiation dose. Synthetic CT (sCT) from non-attenuation corrected (NAC) PET offers a CT-less alternative, but training robust models requires multi-site data that may be difficult to share under privacy regulations. We aim to enable PET-based sCT training across sites without exposing data. METHODS: We built a federated learning (FL) framework and trained two sCT generators–a paired conditional GAN and a CycleGAN. Models were pretrained on a single-site cohort (Site 1, n = 425) and fine-tuned via FL using additional data from Site 1 (n = 25) and a second site with different scanners and reconstruction parameters (Site 2, n = 25). Performance was assessed on an internal hold-out set (Site 1, n = 91) and two external cohorts (Sites 3,4; n = 11, 10) using region-wise relative mean error (rME) of SUV in AC PET. RESULTS: Both models produced anatomically plausible sCT and AC PET with low errors when test data matched training distributions. FL fine-tuning improved robustness under distribution shift at Site 3, reducing errors across most regions, while maintaining comparable performance at Site 4 where protocols resembled the pretraining site. CONCLUSION: Multi-site FL is a feasible path to increase the generalizability of PET-based sCT while preserving data privacy. The proposed framework offers a practical template for training and deploying CT-less AC models across heterogeneous clinical environments.