[Paper Review] DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models

Tiange Xiang, Mahmut Yurt+3 more|arXiv (Cornell University)|Feb 6, 2023

Advanced Neuroimaging Techniques and Applications17 citations

TL;DR

DDM2 introduces a three-stage self-supervised diffusion-based framework for denoising diffusion MRI, representing noisy inputs as intermediate diffusion states and achieving state-of-the-art denoising without ground-truth data.

ABSTRACT

Magnetic resonance imaging (MRI) is a common and life-saving medical imaging technique. However, acquiring high signal-to-noise ratio MRI scans requires long scan times, resulting in increased costs and patient discomfort, and decreased throughput. Thus, there is great interest in denoising MRI scans, especially for the subtype of diffusion MRI scans that are severely SNR-limited. While most prior MRI denoising methods are supervised in nature, acquiring supervised training datasets for the multitude of anatomies, MRI scanners, and scan parameters proves impractical. Here, we propose Denoising Diffusion Models for Denoising Diffusion MRI (DDM$^2$), a self-supervised denoising method for MRI denoising using diffusion denoising generative models. Our three-stage framework integrates statistic-based denoising theory into diffusion models and performs denoising through conditional generation. During inference, we represent input noisy measurements as a sample from an intermediate posterior distribution within the diffusion Markov chain. We conduct experiments on 4 real-world in-vivo diffusion MRI datasets and show that our DDM$^2$ demonstrates superior denoising performances ascertained with clinically-relevant visual qualitative and quantitative metrics.

Motivation & Objective

• Motivate the need for effective denoising of diffusion MRI under diverse anatomies, scanners, and protocols without paired ground-truth data.
• Develop a self-supervised denoising framework that integrates statistical denoising with diffusion models.
• Enable denoising by representing noisy inputs as samples from intermediate states in a diffusion Markov chain.
• Demonstrate cross-dataset generalizability on real-world diffusion MRI datasets.
• Provide an open-source implementation for reproducibility and clinical relevance.

Proposed method

• Stage I learns a slice-to-slice denoising function to estimate an initial noise distribution from limited input volumes.
• Stage II fits a Gaussian noise model to the estimated residuals and matches its standard deviation to a diffusion schedule, locating the input as an intermediate diffusion state.
• Stage III trains a diffusion-based reverse process with two key modifications (noise shuffle and J-Invariance loss) to generate clean approximations without ground-truth supervision.

Experimental results

Research questions

• RQ1 How can unsupervised diffusion models be conditioned on noisy diffusion MRI measurements to produce denoised outputs without paired clean images?
• RQ2 What is the best way to learn and align a data-driven noise model with a diffusion sampling schedule to identify an intermediate state for conditional generation?
• RQ3 Do three-stage self-supervised diffusion denoisers outperform existing supervised and self-supervised MRI denoising methods across diverse datasets?

Key findings

• DDM2 achieves superior denoising performance versus state-of-the-art methods on four real-world diffusion MRI datasets.
• Compared to competing methods, DDM2 yields statistically significant SNR and CNR improvements across evaluated data.
• Compared to the strongest score-based method, DDM2 improves SNR and CNR by averages of 0.95 and 0.93 points respectively.
• DDM2 provides 5–10x faster inference than standard DDPMs due to intermediate-state sampling, while maintaining denoising quality.
• Qualitative results show better restoration of anatomical details with no observed new hallucinations by neuroradiologists.