Deep learning-based intraluminal gas modeling for anatomically accurate synthetic CT in MRI-based radiation therapy.
Authors
Affiliations (3)
Affiliations (3)
- Faculty of Physical and Mathematical Sciences, Benemerita Universidad Autonoma de Puebla, Avenida San Claudio y 18 Sur, Colonia San Manuel, Puebla, 72570, Mexico.
- Meritorious Autonomous University of Puebla, Avenida San Claudio y 18 Sur, Colonia San Manuel, Puebla, 72570, Mexico.
- Radiation Oncology, Harvard Medical School, 75 Francis St., Boston, Massachusetts, 02115-6027, United States.
Abstract
Stochastic bowel and rectal gas complicates MRI/CT deformable image registration (DIR) for synthetic CT (sCT) generation, requiring manual corrections. We propose a DIR-free, two-stage deep learning framework to improve intraluminal gas definition in sCT, streamlining MRI-only simulation and enhancing dosimetric reliability.

Methods: We developed a two-stage GAN framework for sCT generation. First, a CycleGAN or pix2pix model converted MRIs into segmented map images (SMI); second, a pix2pix conditional GAN generated the final sCT. Ground-truth gas cavities were defined using CNN-based gastrointestinal autosegmentation, manual verification, and a 20% intensity threshold. Trained on a 60-patient dataset, the framework was compared to a single-stage MRI-to-sCT model via Dice-Sørensen coefficient and Mean Absolute Error (MAE). Dosimetric accuracy was assessed by recalculating prostate VMAT plans (36.25 Gy/5 fractions) on synthetic CTs.

Results: The gas cavity Dice-Sørensen coefficient for the two-stage method and the single-stage method were 0.63 ± 0.11 and 0.04 ± 0.03 for pix2pix, while for CycleGAN, they were 0.67 ± 0.07 and 0.57 ± 0.11, respectively. The global MAE for the standard single-stage method and the two-stage method for pix2pix were 86±20 HU and 94±20 HU, while for CycleGAN, they were 105±21 HU and 103±22 HU, respectively. Dosimetric analysis demonstrated excellent agreement for the planning target volume (PTV D95%), with mean differences of 0.04 ± 0.08 cGy and -0.89 ± 1.98 cGy for the one-stage and two-stage models, respectively. Both models exhibited negligible deviations across all assessed organ-at-risk (OAR) volumetric metrics, including D0.1cc, D1cc, and D15cc.

Conclusions: The two-stage framework significantly enhances sCT intraluminal gas accuracy, improving cavity definition nearly threefold over single-stage models while maintaining tissue fidelity. This overcomes a major MRI-only simulation bottleneck. Future studies will prospectively evaluate its clinical impact on online adaptive radiotherapy.