Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement

TL;DR

Di-Fusion is a self-supervised diffusion-based denoising framework for diffusion MRI that introduces a Fusion forward process and a Di- noise distribution to capture real-world noise, enabling stable, single-stage training without clean data. By training only the last $T_c$ diffusion steps under a $\,\mathcal{J}$-Invariance objective and employing Run-Walk accelerated sampling with adaptive termination controlled by CSNR, the method achieves iterative, controllable refinement during inference. Across real and simulated datasets, Di-Fusion yields state-of-the-art performance in microstructure modeling, tractography, and diffusion-signal estimates while preserving anatomical fidelity. The work offers practical clinical impact by improving denoising performance under high noise without requiring extra noise models or paired clean data, with code available at the provided GitHub repository $https://github.com/FouierL/Di-Fusion$.

Abstract

Magnetic Resonance Imaging (MRI), including diffusion MRI (dMRI), serves as a ``microscope'' for anatomical structures and routinely mitigates the influence of low signal-to-noise ratio scans by compromising temporal or spatial resolution. However, these compromises fail to meet clinical demands for both efficiency and precision. Consequently, denoising is a vital preprocessing step, particularly for dMRI, where clean data is unavailable. In this paper, we introduce Di-Fusion, a fully self-supervised denoising method that leverages the latter diffusion steps and an adaptive sampling process. Unlike previous approaches, our single-stage framework achieves efficient and stable training without extra noise model training and offers adaptive and controllable results in the sampling process. Our thorough experiments on real and simulated data demonstrate that Di-Fusion achieves state-of-the-art performance in microstructure modeling, tractography tracking, and other downstream tasks. Code is available at https://github.com/FouierL/Di-Fusion.

Figures (32)

• Figure 1: (a) Fusion process (Section \ref{['Fusion process']}) aligns $\left\{ {{{\bar{x}}_t}} \right\}_1^{{T}}$ to $\left\{ {{x_t}} \right\}_1^{{T}}$ and avoids drift ("Drift" means drifted results, "Final" means the denoised version of "Target"); (b) Training the latter diffusion steps (Section \ref{['Intuition of conditional training']}) imposes restrictions on the generation ability of diffusion models and decreases uncertainty; (c) Run-Walk accelerated sampling (Section \ref{['Run-Walk accelerated sampling']}) accelerates the entire sampling process.
• Figure 2: Overview of our single-stage Di-Fusion. The training process does not involve any extra model training apart from ${\cal F}_\theta$, and the sampling process offers adaptive and controllable results.
• Figure 3: Density map of FBC projected on the streamlines of the OR bundles. The numbers in parentheses represent the number of streamlines. Di-Fusion generates the minimal number of streamlines while maintaining high FBCs (consider "Noisy_filtering" as references for high FBCs).
• Figure 4: Scatter plots of the microstructure model predictions against input data. The top-left of each plot shows the quantitative $R^2$ metric computed from each model fit on the corresponding data. Our data points are more concentrated (higher $R^2$).
• Figure 5: Qualitative results. "OURS" results are obtained when ${\mathcal{CSNR}}=0.040$. The area indicated by the red arrow does not appear in "OURS", indicating that Di-Fusion does not remove structural information during denoising.