Abstract
Cone-beamcomputedtomography(CBCT)isroutinelyused for patient positioning in image-guided radiotherapy. However, its poor imagequalitypreventsitsdirectuseinradiotherapyreplanning.Recent deeplearningmethodsattempttotranslateCBCTtoCTvolumes,however they are based on generative models such as GANs and Diffusion, which either yield suboptimal image quality or are computationally expensive.Inthispaper,weshowthatalight-weightmodelissufficientfor generating CT volumes from CBCT inputs in a single step. Our model outperforms Diffusion and GAN models in terms of voxel reconstruction accuracy while running over 20× faster than score-based diffusion methods(thepreviousstateoftheart),offeringanefficientandreliable solution for adaptive clinical workflows. Our architecture is optimized via a decoupled two-stage paradigm: a lightweight convolutional networkequippedwithFeature-wiseLinearModulation(FiLM)firstaligns cross-modalitylatentfeatures,followedbyatargeted,multi-lossdecoder fine-tuning stage that restores fine image details and ensures spatial fidelity. To overcome the scarcity of clinical pairs and augment training data,weintegrateaphysics-basedsimulationprocessofsyntheticCBCT.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/SASHIMI_048.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=iQYgJsqO3x
BibTex
@InProceedings{LonPau_ALatent_MICCAISAT2026,
author = { Londres, Paul AND Bibault, Jean-Emmanuel AND Lepetit, Vincent AND Paragios, Nikos},
title = { { A Latent Single-Step Network is Sufficient for Accurate CBCT-CT Translation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}