Metal implants such as dental fillings, spinal fixation systems, hip prostheses, and coils cause severe streaking and shading artifacts in CT because of beam hardening, photon starvation, scatter, nonlinear partial-volume effects, and inconsistencies in the projection data. These artifacts can obscure anatomy, bias attenuation values, and reduce confidence in downstream tasks such as segmentation, treatment planning, and image-guided intervention. Classical methods such as linear interpolation and normalized metal artifact reduction (NMAR) remain useful, but they often blur detail or introduce secondary artifacts around complex, dense implants (Gjesteby et al., 2016 Meyer et al., 2010).
Recent deep-learning MAR methods have improved image quality in the sinogram domain, image domain, and dual-domain setting. In particular, diffusion-based methods now define a strong research direction learning richer image priors and often preserve structure better than earlier CNN or GAN approaches. Recent works highlight progress from data-consistent dual-domain networks such as InDuDoNet+ (Wang et al., 2023) to conditional diffusion models such as DCDiff (Shen et al., 2024), as well as the emergence of scalable latent diffusion frameworks for medical imaging (Rombach et al., 2022), including MAISI and MedLoRD (Guo et al., 2024). These developments suggest that conditioned latent diffusion is a promising foundation for CT metal artifact suppression because it combines powerful generative priors with substantially lower computational cost than pixel-space diffusion

The main objective is to develop and validate a generative AI method for CT metal artifact suppression based on conditioned latent diffusion.
Specific objectives are to: (1) design a conditioning strategy that combines the corrupted CT image with artifact-aware priors (2) train and compare different latent diffusion variants for artifact suppression (3) benchmark the proposed method against classical and modern MAR baselines and (4) quantify anatomical fidelity, artifact suppression, and robustness on real clinical data.

The developments will be performed in close collaboration with the university hospital UZ Brussel. The research will mainly address the exploration of the novel generative AI methodologies for metal artifact supression from 3D CT volumes. This thesis will expose the student to a broad set of tasks and competences, covering experimental work and algorithm development and testing. The following tasks can be distinguished:
•
Literature review on CT metal artifact reduction, diffusion models, latent diffusion, and weakly supervised medical image restoration.
•
Implementation of preprocessing, baseline MAR methods, and the proposed conditioned latent diffusion model in Python (PyTorch and MONAI).
•
Training and validation on simulated and real CT data, including ablation studies on different conditioning strategies.
•
Comparison with state-of-the-art baselines using quantitative metrics and targeted visual analysis around metal implants.
•
Interpretation of failure cases, discussion of clinical relevance, and full thesis writing.