Predicting Synthetic CT from Ultrasound: A Cycle-Consistent Diffusion Network for Prostate High-Dose-Rate Brachytherapy
Abstract
Purpose
Ultrasound-guided high-dose-rate (UGHDR) brachytherapy for prostate cancer depends on real-time transrectal ultrasound (TRUS) imaging for catheter guidance. However, the limited ability of TRUS to depict critical bony anatomy, such as the pubic arch, poses challenges for safe and accurate needle insertion. This constraint has motivated the development of US-derived synthetic CT (sCT) to improve intraoperative anatomical visualization while preserving real-time imaging capabilities.
Methods
A Cycle Diffusion Network is developed for US-to-CT synthesis to generate anatomically consistent synthetic CT images from US. The framework is built upon diffusion models due to their stable training behavior and strong capability in reconstructing high-fidelity anatomical structures from degraded inputs. To ensure accurate cross-modality correspondence, the proposed network employs a dual forward–reverse denoising architecture that explicitly models bidirectional US-to-CT and CT-to-US transformations. Two conditional diffusion branches are used, each learning modality-specific mappings while sharing complementary structural information. In addition, a diffusion-based cycle-consistency constraint is incorporated to enforce geometric alignment between synthesized CT and the corresponding US, reducing artifacts caused by US noise and appearance discrepancies.
Results
The proposed Cycle Diffusion Network was evaluated against Pix2Pix and CycleGAN for US -to-CT synthesis. Qualitative results demonstrate that the proposed method generates synthetic CT images with improved anatomical consistency, particularly in reconstructing pelvic bony structures while better preserving soft tissue morphology around the prostate. Quantitative comparisons further confirm the superiority of our method, achieving a normalized cross-correlation (NCC) of 0.98 ± 0.01 and a normalized mutual information (NMI) of 0.46 ± 0.13, substantially outperforming the competing methods.
Conclusion
The proposed Cycle Diffusion Network enables structure-aware US-to-CT synthesis, outperforming GAN-based methods by preserving bone geometry and tissue boundaries, demonstrating strong potential for reliable intraoperative guidance in prostate HDR brachytherapy.