Diffusion transformer generalizes proton Monte Carlo denoising, but simple filtering remains competitive

A head-and-neck-trained diffusion transformer generalized proton dose denoising across three unseen sites, but optimized Gaussian filtering matched or outperformed several global metrics.

KEY POINTS

  • The diffusion-transformer model was trained on 80 head and neck IMPT patients and internally tested on another 10. Cross-site generalizability was then evaluated without retraining in 30 independent patients with lung, breast, and prostate cancer, 10 from each disease site.
  • Training inputs consisted only of CT information and a low-statistics Monte Carlo dose distribution generated with 1/100 of the particles used for the high-statistics reference calculation. Structure contours were not supplied as predictive inputs, making the cross-site test less dependent on disease-specific anatomy.
  • Across the four test cohorts, whole-body mean absolute error ranged from 0.08 to 0.20 Gy(RBE). Mean whole-body 3%/2 mm gamma passing rates ranged from 95.8% to 97.6%, despite lung, breast, and prostate being unseen during training.
  • Cross-site performance was not uniform in clinically important regions. CTV mean absolute error was 0.14 Gy(RBE) for head and neck, but increased to 1.7 Gy for lung, 1.2 Gy for breast, and 1.0 Gy for prostate, reflecting the difficulty of out-of-distribution target-dose reconstruction.
  • The most important comparator substantially tempers the AI result. An optimized Gaussian filter produced lower whole-body MAE than the diffusion transformer in head and neck, lung, and breast, and the same rounded MAE in prostate; its whole-body gamma rate was also higher in head and neck, lung, and prostate and identical in breast.
  • The learned model did preserve some target metrics better: in the head and neck cohort, CTV D98 deviation was 0.8% with the diffusion transformer versus 1.2% with Gaussian filtering, although Gaussian filtering performed better for CTV D2 deviation (0.7% vs 1.3%).
  • Low-statistics Monte Carlo calculation required approximately 1 minute, followed by 40–120 seconds for diffusion-transformer inference, compared with approximately 10 minutes for high-statistics Monte Carlo in representative head and neck cases. The dataset was single-institutional, range-shifter cases were excluded, and no optimized convolutional neural-network comparator was tested.

CLINICAL TAKEAWAY

The study shows that a proton Monte Carlo denoising model trained at one heterogeneous disease site can transfer to lung, breast, and prostate without retraining, which is attractive for time-sensitive adaptive workflows. But the diffusion transformer was not uniformly superior to a simple optimized Gaussian filter, so the value of the added model complexity remains unproven and requires stronger benchmarking before clinical deployment.

SOURCE

Physics and Imaging in Radiation Oncology

Browse more research Suggest a correction