The pursuit of accurate MRI-to-CT synthesis is paramount for enabling MR-only pelvic workflows, offering CT-like bone detail without ionizing radiation. This research benchmarks recently proposed drifting models for CT synthesis against a suite of established methods, including UNet, VAE, WGAN-GP, PPFM, and various diffusion models (FastDDPM, DDIM, DDPM).
Drifting Models Emerge as State-of-the-Art for Pelvic CT Synthesis
Across two distinct datasets, the drifting model demonstrated a clear performance advantage. It achieved superior image fidelity and structural consistency, evidenced by higher SSIM and PSNR, and lower RMSE compared to strong diffusion baselines and conventional CNN-, VAE-, GAN-, and PPFM-based approaches. Qualitative assessments highlighted sharper cortical bone edges, improved geometric depiction of critical anatomical structures like the sacrum and femoral heads, and a marked reduction in artifacts and over-smoothing, particularly at challenging bone-air-soft tissue interfaces. These findings position drifting models for CT synthesis as a significant leap forward.
Unprecedented Efficiency for Clinical Translation
Beyond raw image quality, the drifting model achieves these gains with a remarkable efficiency profile. It requires only a single inference step, completing synthesis in milliseconds. This contrasts sharply with iterative diffusion sampling methods, offering a more favorable accuracy-efficiency trade-off. The speed and quality make drifting models a highly promising direction for fast, high-quality pelvic synthetic CT generation from MRI, directly addressing the need for rapid and reliable imaging in clinical settings such as radiotherapy planning and PET/MR attenuation correction.