[Paper Review] DDMM-Synth: A Denoising Diffusion Model for Cross-modal Medical Image Synthesis with Sparse-view Measurement Embedding

Xiaoyue Li, Kai Shang+2 more|arXiv (Cornell University)|Mar 28, 2023

Medical Imaging Techniques and Applications9 citations

TL;DR

DDMM-Synth uses an MRI-guided diffusion model with a range-null space decomposition to synthesize CT from MRI while incorporating sparse-view CT measurements, improving data consistency and robustness to noise.

ABSTRACT

Reducing the radiation dose in computed tomography (CT) is important to mitigate radiation-induced risks. One option is to employ a well-trained model to compensate for incomplete information and map sparse-view measurements to the CT reconstruction. However, reconstruction from sparsely sampled measurements is insufficient to uniquely characterize an object in CT, and a learned prior model may be inadequate for unencountered cases. Medical modal translation from magnetic resonance imaging (MRI) to CT is an alternative but may introduce incorrect information into the synthesized CT images in addition to the fact that there exists no explicit transformation describing their relationship. To address these issues, we propose a novel framework called the denoising diffusion model for medical image synthesis (DDMM-Synth) to close the performance gaps described above. This framework combines an MRI-guided diffusion model with a new CT measurement embedding reverse sampling scheme. Specifically, the null-space content of the one-step denoising result is refined by the MRI-guided data distribution prior, and its range-space component derived from an explicit operator matrix and the sparse-view CT measurements is directly integrated into the inference stage. DDMM-Synth can adjust the projection number of CT a posteriori for a particular clinical application and its modified version can even improve the results significantly for noisy cases. Our results show that DDMM-Synth outperforms other state-of-the-art supervised-learning-based baselines under fair experimental conditions.

Motivation & Objective

• Motivate reducing CT radiation by enabling high-fidelity CT reconstruction from sparse measurements and cross-modal MRI information.
• Develop a diffusion-model framework that leverages MRI guidance and sparse-view CT priors for MRI-to-CT synthesis.
• Enable flexible adjustment to measurement settings and robustness to noisy measurements without retraining.
• Assess whether integrating MRI guidance and CT measurement embedding yields superior synthesis quality compared to state-of-the-art baselines.

Proposed method

• Conditional diffusion model trained with MRI guidance.
• Null-space range-null space decomposition to insert sparse-view CT information during reverse sampling.
• CT reconstruction consistency enforced via A†y and measurements y in the reverse process (Equations 5–7).
• Noise handling in diffusion steps with gamma_t and phi_t to account for measurement noise (Equations 7–13).
• DDIM sampling to accelerate reverse process (T = 2000 to 100 steps).
• A single framework adaptable to different linear measurement mappings without retraining.

Experimental results

Research questions

• RQ1 Can an MRI-guided diffusion model, combined with sparse-view CT measurements, produce higher-fidelity CT images than prior MRI-to-CT translation or purely data-driven reconstruction methods?
• RQ2 Does the range-null space measurement embedding improve data consistency and structural fidelity in synthesized CT under sparse-view and noisy scenarios?
• RQ3 How does DDMM-Synth perform relative to state-of-the-art supervised baselines on multi-modal pelvic MRI-CT (Gold Atlas) and BRATS2018 datasets?
• RQ4 Is the method robust to different projection counts and to measurement noise without additional model training?

Key findings

• DDMM-Synth achieves higher PSNR/SSIM than several baselines on Gold Atlas and BRATS2018 under varying projection counts.
• With fewer projections (e.g., 20 vs 23) DDMM-Synth can outperform some methods that use more projections, indicating strong data-consistency with sparse views.
• DDMM-Synth-noise variant significantly improves performance in noisy acquisitions, as evidenced by FID reductions.
• Ablation studies show MRI guidance plus sparse-view CT embedding yields more faithful CT reconstructions than variants lacking either component.
• Diffusion-based synthesis provides higher anatomical fidelity and fewer artifacts in ROIs compared to non-diffusion baselines.