Abstract
Recent wavelet-domain diffusion models improve synthetic CT generation by modeling images in multi-scale frequency spaces. However, most methods still process low- and high-frequency components jointly or use high-frequency information as auxiliary guidance. This design can be effective when source and target modalities are relatively aligned, but it offers limited transparency about frequency-specific errors and uncertainty. This limitation becomes more evident in large-gap translation such as MRI-to-CT, where low-frequency soft-tissue anatomy is partially recoverable from MRI, but CT-specific high-frequency structures such as bone edges, tissue boundaries, and local textures are weakly constrained.
We propose a frequency-decoupled diffusion framework that treats high-frequency CT texture as an independent stochastic generative target. The target CT is decomposed into low-frequency anatomical and high-frequency texture components. An Anatomical Structure Modeling branch preserves global anatomical structure, while a Tissue Texture Modeling branch independently synthesizes stochastic high-frequency texture and enables sampling-based uncertainty estimation. Experiments on SynthRAD2025 show that the proposed method achieves substantial improvements on MRI-to-CT synthesis, obtaining 24.1 dB PSNR and 0.84 SSIM, while remaining comparable on CBCT-to-CT synthesis with 26.6 dB PSNR and 0.87 SSIM. These results suggest that explicit stochastic high-frequency modeling is especially beneficial when the source–target modality gap is large, and the resulting uncertainty maps provide interpretable cues for ambiguous texture reconstruction in radiotherapy-oriented sCT generation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MWM_114.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=47YJorOYnV
BibTex
@InProceedings{ZhaChe_Stochastic_MICCAISAT2026,
author = { Zhang, Chen AND Ling, Sai Ho AND Shimizu, Akinobu},
title = { { Stochastic High-Frequency Modeling for Uncertainty-Aware Synthetic CT Generation in Radiotherapy } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}
