Leveraging Image Editing Foundation Models for Data-Efficient CT Metal Artifact Reduction

A research team from Turkey has published a novel method that adapts a general-purpose vision-language diffusion foundation model to solve a critical medical imaging problem: reducing metal artifacts in CT scans. These artifacts, caused by high-attenuation implants like hip replacements or dental fillings, severely degrade image quality and obscure anatomy. The researchers reframed the problem as an in-context reasoning task for the foundation model, using parameter-efficient Low-Rank Adaptation (LoRA) for fine-tuning. This approach leverages the model's rich visual priors, allowing it to achieve effective artifact suppression with a dramatically reduced dataset of just 16 to 128 paired training examples—a reduction of two orders of magnitude compared to standard deep learning methods.

Crucially, the team discovered that domain adaptation is essential to prevent the model from misinterpreting the artifacts. Without proper grounding, the unadapted foundation model would hallucinate, interpreting streak artifacts as erroneous real-world objects like waffles or petri dishes. To solve this, they developed a multi-reference conditioning strategy. This technique provides the model with clean anatomical exemplars from unrelated subjects alongside the corrupted input scan, enabling it to use category-specific context to infer the correct, uncorrupted anatomy. Extensive evaluation on the AAPM CT-MAR benchmark shows the method achieves state-of-the-art performance on both perceptual and radiological-feature metrics, establishing a scalable, data-efficient paradigm for medical image reconstruction.