# DDM$^2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models

**Authors:** Tiange Xiang, Mahmut Yurt, Ali B Syed, Kawin Setsompop, Akshay Chaudhari  
**arXiv:** [2302.03018](https://arxiv.org/abs/2302.03018)  
**Date:** 2023-02-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2302.03018v1/

## Key points

- Magnetic resonance imaging (MRI) is a widely used medical imaging technique
- High-quality MRI scans with high signal-to-noise ratio (SNR) require long scan times
- Long scan times lead to increased costs, patient discomfort, and reduced throughput
- Denoising methods are needed to improve the quality of MRI scans, especially for diffusion MRI scans
- The authors propose a self-supervised denoising method called DDM$^2$
- DDM$^2$ leverages diffusion denoising generative models and integrates statistic-based denoising theory into diffusion models
- The DDM$^2$ framework consists of three stages: conditional generation, integration of statistic-based denoising theory into diffusion models, and denoising through conditional generation
- DDM$^2$ outperforms existing methods in terms of denoising performance according to experiments conducted on real-world in-vivo diffusion MRI datasets
- DDM$^2$ achieves superior denoising performance based on clinically-relevant visual qualitative and quantitative metrics
- DDM$^2$ presents a promising self-supervised approach for denoising diffusion MRI scans
- Integration of diffusion models with statistic-based denoising theory and leveraging conditional generation techniques contribute to the superior performance of DDM$^2$
- DDM$^2$ has significant implications for improving the efficiency and quality of MRI scans, benefiting patients and healthcare providers.

## Summary

Magnetic resonance imaging (MRI) is a widely used medical imaging technique that plays a crucial role in diagnosing and treating various conditions. However, obtaining high-quality MRI scans with a high signal-to-noise ratio (SNR) typically requires long scan times, leading to increased costs, patient discomfort, and reduced throughput. Consequently, there is a growing need for denoising methods to improve the quality of MRI scans, particularly for diffusion MRI scans that are severely SNR-limited. To address this challenge, the authors propose a self-supervised denoising method called Denoising Diffusion Models for Denoising Diffusion MRI (DDM$^2$). This approach leverages diffusion denoising generative models and integrates statistic-based denoising theory into diffusion models. The DDM$^2$ framework consists of three stages: conditional generation, integration of statistic-based denoising theory into diffusion models, and denoising through conditional generation. During inference, the input noisy measurements are represented as samples from an intermediate posterior distribution within the diffusion Markov chain. To evaluate the performance of DDM$^2$, experiments were conducted using four real-world in-vivo diffusion MRI datasets. The results demonstrate that DDM$^2$ outperforms existing methods in terms of denoising performance. This superiority was confirmed through clinically-relevant visual qualitative and quantitative metrics. In conclusion, DDM$^2$ presents a promising self-supervised approach for denoising diffusion MRI scans. By integrating diffusion models with statistic-based denoising theory and leveraging conditional generation techniques, DDM$^2$ achieves superior denoising performance compared to existing methods. This advancement has significant implications for improving the efficiency and quality of MRI scans which ultimately benefits both patients and healthcare providers alike.