Abstract
Metal artifacts in computed tomography (CT) significantly degrade anatomical visibility and can compromise downstream image analysis tasks, including segmentation, registration, implant migration assessment, and image-guided surgical planning. Many effective metal artifact reduction methods use projection-domain or dual-domain information, but raw sinograms are often unavailable in retrospective or public clinical CT datasets. In this work, we present the first 3D imagedomain framework for large-volume CT metal artifact suppression using conditional latent diffusion models (LDM). First, a VQ-VAE-GAN compresses CT volumes to latent representations. Secondly, a 3D conditioned diffusion U-Net operating in a latent space on crops of 448 × 448 × 256 suppresses CT metal artifacts by synthesizing non-artifacted versions of the input. We compare two conditional models: an anatomy conditioned LDM and an anatomy-metadata conditioned LDM with extra information on anatomical region, laterality, and implant material. Training is performed on paired synthetic data generated by an anatomy-aware metal artifact generation pipeline that inserts implant masks into clean CT volumes, simulates polychromatic projection artifacts, and reconstructs artifact-corrupted volumes with paired non-artifacted image. On 50 test volumes, the anatomy-metadata conditioned model achieved the strongest structural and perceptual metrics (RMSE 0.021, PSNR 39.430 dB, SSIM 0.992, LPIPS 0.028), while anatomy-only conditioning showed slightly stronger visual artifact suppression. On postoperative real CTs with metal artifacts, the method successfully reduced visible artifacts in qualitative examples. The results support large-volume 3D image-domain LDMs as a promising direction for CT metal artifact suppression.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_025.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/profile?id=~Xabier_Moreno_Casado1
BibTex
@InProceedings{MorXab_LargeVolume_MICCAISAT2026,
author = { Moreno Casado, Xabier AND Vandemeulebroucke, Jef AND Ceranka, Jakub},
title = { { Large-Volume Conditioned 3D Latent Diffusion Models for CT Metal Artifact Suppression } },
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}
}