2255 - Diffusion-Synthetic MRI for Privacy-Preserving Hybrid Training of Brain Tumor Auto-Segmentation in Radiotherapy
Presenter(s)
Y. J. Wang1, T. L. Yu2, M. F. Lin3, W. L. Tseng2, and F. R. Xiao3; 1Department of Radiation Oncology, Fu Jen Catholic University Hospital, Fu Jen Catholic University, New Taipei City, Taiwan, 2Fu Jen Catholic University Hospital, New Taipei City, Taiwan, 3Institute of Medical Device and Imaging ,National Taiwan University College of Medicine, Taipei, Taiwan
Purpose/Objective(s):
Multi-institution development of MRI-based auto-contouring can reduce radiotherapy workload, but is limited by privacy constraints and scarce expert labels. We tested the hypothesis that diffusion-generated synthetic contrast-enhanced (CE) T1-weighted brain MRI, combined with limited real data, can yield robust brain tumor segmentation with physician-rated clinical usability while reducing the need to share patient imaging.Materials/Methods:
A medical denoising diffusion probabilistic model (Med-DDPM) generated synthetic CE T1-weighted brain MRI. Three 3D segmentation networks (nnU-Net, U-Mamba, LightM-UNet) were trained under synthetic-only training and hybrid real/synthetic settings. Real data comprised 20 cases (R20) assembled across two institutions (tumor masks from one site and clinical MRI from another). Hybrid settings added 10/100/1000 synthetic cases (R20+S10/R20+S100/R20+S1000). Endpoints included overlap accuracy (Dice, IoU), false-positive burden, and clinical usability. Clinical validation used an independent retrospective cohort of 40 CE T1 MRI cases not used in training; segmentations were reviewed by radiation oncologists and neurosurgeons for omission rate, edit burden, and processing time.Results: In synthetic-only training, U-Mamba achieved the best overlap performance (Dice 0.9634; IoU 0.9312) versus nnU-Net (Dice 0.9592; IoU 0.9234) and LightM-UNet (Dice 0.9334; IoU 0.8785). In hybrid training, moderate synthesis (R20+S100) provided stable generalization with a balanced precision profile, whereas heavy synthesis (R20+S1000) increased false positives despite high overlap. In clinical validation using U-Mamba, mean omission rate was 8.57% (median 0%); >80% of cases required only minor edits for clinical use. Average AI processing time was ~3 minutes/case versus 20-60 minutes for manual contouring. In a highly complex case with >60 lesions, most lesions were detected, with occasional vascular false positives requiring physician review.
Conclusion: Diffusion-generated synthetic MRI can augment small real datasets to support privacy-preserving model development for clinically efficient brain tumor segmentation in radiotherapy workflows. Moderate synthetic augmentation appeared to best balance accuracy and false positives. Prospective multi-center validation and predefined physician review safeguards remain necessary prior to routine clinical deployment.