A strategy for simulation-driven CT metal artifact reduction toward improving network generalizability.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Physics

Background:

• Existing deep learning methods for computed tomography (CT) metal artifact reduction (MAR) often lack explicit physical modeling, leading to issues like hallucination and anatomical distortion.
• These data-driven approaches require extensive paired datasets, limiting their scalability and generalizability across different imaging scenarios.

Purpose of the Study:

• To develop a novel self-supervised framework for CT MAR that integrates physics-informed deep learning to improve artifact reduction.
• The framework aims to enhance scalability, reduce hallucination, and preserve structural fidelity in CT images by combining a multi-layer perceptron (MLP) with a latent diffusion model (LDM).

Main Methods:

• A lightweight multi-layer perceptron (MLP) performs physics-driven polynomial correction for beam-hardening, incorporating sinogram consistency for adaptation.
• The MLP's learned parameters simulate artifact-contaminated images from artifact-free scans, creating pseudo-paired data for self-supervised training of a conditional latent diffusion model (LDM).

Main Results:

• The proposed framework demonstrates superior metal artifact reduction and structural preservation compared to state-of-the-art techniques on both synthetic and real clinical datasets.
• Operating in a low-dimensional latent space, the latent diffusion model (LDM) significantly reduces inference time while maintaining high-quality image reconstruction.

Conclusions:

• A self-supervised CT MAR framework combining MLP-based beam-hardening correction and a conditional LDM was successfully developed.
• The framework enables fully self-supervised training without paired data, offering robust artifact suppression and structural preservation, outperforming existing methods.