Deep learning-based intraluminal gas modeling for anatomically accurate synthetic CT in MRI-based radiation therapy

Purpose.Accurate modeling of intraluminal gas in synthetic computed tomography (sCT) is often compromised by the stochastic nature of bowel and rectal gas, which complicates magnetic resonance imaging (MRI)/CT deformable image registration (DIR) and necessitates time-consuming manual corrections. Here, we propose a novel, DIR-free, two-stage deep learning framework designed to improve the definition of intraluminal gas in sCT. By circumventing registration-based errors, this method aims to streamline MRI-only simulation and enhance the dosimetric reliability of sCT images in MRI-based radiotherapy.Methods.sCT generation was implemented using a two-stage generative adversarial network (GAN) framework. In the first stage, a CycleGAN or pix2pix model converted MRI inputs into segmented map images (SMI); in the second stage, a conditional GAN (pix2pix) transformed the SMI into the final sCT. Ground-truth intraluminal gas cavities were defined in the following order: gastrointestinal contours were generated via CNN-based autosegmentation, manually verified, and subjected to a 20% intensity threshold. The framework was trained and validated using a publicly available 60-patient dataset. Performance was evaluated using the Dice-Sørensen coefficient for gas definition and mean absolute error (MAE) for global and tissue-specific Hounsfield unit (HU) accuracy, comparing the two-stage framework against a direct single-stage MRI-to-sCT model. To evaluate dosimetric accuracy, prostate volumetric modulated arc therapy plans (36.25 Gy in 5 fractions) were optimized on five reference bulk-density sCT datasets and recalculated on both one-stage and two-stage sCT using a clinical treatment planning system.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 volumetric metrics, including D0.1cc, D1cc, and D15cc.Conclusions.The proposed two-stage framework significantly enhances the anatomical accuracy of intraluminal gas in sCT images, achieving a nearly threefold improvement in gas cavity definition over standard direct-conversion models. This method maintains high fidelity for bone and soft tissue while addressing a critical bottleneck in MRI-only simulation. Future studies will focus on the prospective evaluation of this framework to quantify its impact on dosimetric accuracy and treatment efficiency in online adaptive radiotherapy workflows.