[Paper Review] Cycle-guided Denoising Diffusion Probability Model for 3D Cross-modality MRI Synthesis

Shaoyan Pan, Chih‐Wei Chang+8 more|arXiv (Cornell University)|Apr 28, 2023

Radiomics and Machine Learning in Medical Imaging18 citations

TL;DR

This paper introduces a Cycle-guided Denoising Diffusion Probability Model (CG-DDPM) that uses two conditional DDPMs exchanging latent noise to synthesize matching 3D MRI modalities across two pulse sequences, improving cross-modality synthesis accuracy and consistency.

ABSTRACT

This study aims to develop a novel Cycle-guided Denoising Diffusion Probability Model (CG-DDPM) for cross-modality MRI synthesis. The CG-DDPM deploys two DDPMs that condition each other to generate synthetic images from two different MRI pulse sequences. The two DDPMs exchange random latent noise in the reverse processes, which helps to regularize both DDPMs and generate matching images in two modalities. This improves image-to-image translation ac-curacy. We evaluated the CG-DDPM quantitatively using mean absolute error (MAE), multi-scale structural similarity index measure (MSSIM), and peak sig-nal-to-noise ratio (PSNR), as well as the network synthesis consistency, on the BraTS2020 dataset. Our proposed method showed high accuracy and reliable consistency for MRI synthesis. In addition, we compared the CG-DDPM with several other state-of-the-art networks and demonstrated statistically significant improvements in the image quality of synthetic MRIs. The proposed method enhances the capability of current multimodal MRI synthesis approaches, which could contribute to more accurate diagnosis and better treatment planning for patients by synthesizing additional MRI modalities.

Motivation & Objective

• Develop a diffusion-based framework for cross-modality MRI synthesis between two pulse sequences.
• Enable mutual regularization between two conditional DDPMs through latent noise exchange.
• Improve synthesis accuracy and modality-consistency for multi-modal MRI generation.
• Evaluate performance against existing methods on a standard dataset to demonstrate improvements.

Proposed method

• Employ two conditional denoising diffusion probability models that generate synthetic images for each MRI modality.
• Exchange random latent noise in the reverse diffusion processes to regularize both models.
• Train the pair with cycle-consistency aims to ensure matching across modalities.
• Quantitatively evaluate using MAE, MSSIM, and PSNR, and assess synthesis consistency on BraTS2020.
• Compare CG-DDPM with several state-of-the-art networks to establish statistical significance of improvements.

Experimental results

Research questions

• RQ1 Can cycle-guided latent noise exchange between two conditional DDPMs improve cross-modality MRI synthesis accuracy?
• RQ2 Does the proposed CG-DDPM achieve higher synthesis consistency between two MRI modalities compared to existing methods?
• RQ3 How does CG-DDPM perform on standard multi-modal MRI datasets in terms of MAE, MSSIM, and PSNR?
• RQ4 Are the improvements statistically significant relative to baseline networks?

Key findings

• CG-DDPM achieves higher synthesis accuracy and reliable modality-consistency on the BraTS2020 dataset.
• The dual DDPM setup with latent noise exchange provides regularization benefits leading to improved image quality.
• Compared to several state-of-the-art networks, CG-DDPM shows statistically significant improvements in MRI synthesis quality.
• The method enhances multimodal MRI synthesis potential, aiding more accurate diagnosis and treatment planning.