KEY POINTS
- This retrospective multicentre study included 2,187 patients with biopsy-proven nasopharyngeal carcinoma from 39 institutions. The federated model was trained on 2,061 patients from 14 centres and externally validated on 126 patients from 25 centres using data from 92 magnetic resonance imaging scanners.
- Contrast-free T1-weighted and T2-weighted images were used to synthesize virtual contrast-enhanced images, with gadolinium-enhanced imaging as the reference. Federated training allowed institutions to collaborate without transferring patient imaging data.
- On external data, the federated model achieved a mean absolute error of 45.85 (95% confidence interval, 43.73–47.96), compared with 58.21 (95% confidence interval, 57.52–58.90) for the average single-institution model—a 21.23% improvement.
- Fine-tuning the federated model at three underperforming institutions reduced mean absolute error from 53.93 to 45.93, representing an average improvement of 14.83%.
- Ten clinicians from eight centres evaluated external cases. Virtual images were rated suitable for diagnosis in 98.98%, staging in 97.96%, and tumor delineation in 92.33%, compared with 99.21%, 97.62%, and 91.27% for real contrast-enhanced images. Clinician accuracy in distinguishing synthetic from real images was 51.27%.
CLINICAL TAKEAWAY
Federated learning may enable contrast-free magnetic resonance imaging support for nasopharyngeal cancer diagnosis, staging, and radiotherapy contouring while preserving institutional data privacy. However, the study was retrospective, lacked prospective task-specific validation, and identified a potential risk of smoothing small infiltrative extensions, so the findings are technically relevant rather than practice-changing.
SOURCE
International Journal of Radiation Oncology, Biology, Physics