BM4D-PC: nonlocal volumetric denoising of principal components of diffusion-weighted MR images

Purpose: Noise in diffusion-weighted MRI (dMRI) is often spatially correlated due to different acquisition and reconstruction strategies, which is not fully accounted for in current denoising strategies. Thus, we propose a novel model-based denoising method for dMRI that effectively accounts for the different noise characteristics of data. Methods: We propose a denoising strategy that incorporates full noise statistics, including the noise power spectral density (PSD), by leveraging the BM4D algorithm. Furthermore, to exploit redundancy across the diffusion MRI dataset, BM4D is applied to principal components (PC) of diffusion-weighted images (DWI) obtained through principal component analysis (PCA) decomposition of the entire DWI dataset, an approach we refer to as BM4D-PC. Importantly, our method also allows for direct estimation of both the noise map and PSD. We evaluated BM4D-PC against four existing state-of-the-art methods using in-silico and in vivo datasets, including high-resolution human and marmoset acquisitions. Results: Overall, BM4D-PC presented the best results for the metrics PSNR, SSIM and RMSE on the in-silico experiments. The in-vivo studies also showed that BM4D-PC dramatically enhanced the image quality of raw DWIs, outperforming existing denoising methods in terms of noise suppression and detail preservation, leading to improved quality of diffusion metrics. Conclusion: The proposed BM4D-PC method demonstrated state-of-the-art denoising results for dMRI using datasets from various acquisition strategies and image resolutions, potentially supporting future advances in neuroscience research.

💡 Research Summary

This paper introduces BM4D‑PC, a novel model‑based denoising framework specifically designed for diffusion‑weighted MRI (dMRI) data that exhibit spatially correlated (colored) noise. Traditional dMRI denoising methods such as MPPCA, NORDIC, Tensor‑MPPCA, and self‑supervised deep learning approaches typically assume independent and identically distributed (i.i.d.) white Gaussian noise. However, many acquisition and reconstruction strategies—partial Fourier, zero‑filling, non‑Cartesian spiral trajectories, and advanced parallel‑imaging reconstructions—produce noise that is spatially correlated and characterized by a non‑flat power spectral density (PSD). Ignoring this PSD leads to sub‑optimal noise suppression and potential bias in downstream diffusion metrics.

BM4D‑PC addresses these limitations in two main stages. First, the entire set of diffusion‑weighted images (DWIs) is reshaped into a matrix whose columns contain vectorized 3‑D volumes. A global principal component analysis (PCA) is performed on this matrix, yielding orthogonal eigenvectors (principal components, PCs) and associated singular values. Because PCA is a unitary transformation, the noise statistics—including spatial correlation and PSD—are preserved across all PCs. The signal energy concentrates in the first few PCs, while the last PCs are dominated by noise only.

Second, each PC volume is denoised independently using the BM4D algorithm, a non‑local block‑matching and 4‑D transform‑domain filtering method originally developed for i.i.d. Gaussian noise. The authors employ the extended version of BM4D that incorporates the estimated noise PSD, allowing accurate computation of transform‑domain noise variances even for colored noise. Block‑matching is performed only on the highest‑SNR PC; the resulting block coordinates are reused for all other PCs, thereby stabilizing the matching process for low‑SNR components. BM4D proceeds in two stages: a hard‑thresholding stage that produces a basic estimate, followed by a Wiener‑filtering stage that refines the estimate using the pre‑denoised image.

A key advantage of BM4D‑PC is its fully automatic estimation of both the spatially varying noise map and the noise PSD directly from the data. The authors exploit the fact that the last few PCs contain almost pure noise; a local standard‑deviation estimator applied within a small 3‑D neighborhood yields a voxel‑wise noise map. After normalizing the PCs by this map, the residual noise becomes stationary, enabling PSD estimation via local 2‑D Fourier transforms on overlapping slice blocks. The final PSD is up‑sampled to match the full 3‑D resolution. The estimation is performed using only the highest b‑value shell, which minimizes signal contamination.

The method was validated on two fronts. In silico, a publicly available noise‑free diffusion phantom (Fiberfox‑based ISMRM 2015 challenge) was corrupted with both white and colored complex Gaussian noise at 1 %, 5 %, and 10 % of the b = 0 signal intensity. Quantitative metrics (PSNR, SSIM, RMSE) showed that BM4D‑PC consistently outperformed MPPCA, NORDIC, Tensor‑MPPCA, and Patch2Self across all noise levels and types. In vivo, high‑resolution human EPI data and ultra‑high‑resolution marmoset data were processed. Visual inspection revealed superior noise suppression while preserving fine anatomical details such as fiber crossings and cortical boundaries. Importantly, diffusion tensor and kurtosis metrics (FA, MD, etc.) derived from BM4D‑PC‑denoised data exhibited reduced bias and variance compared with other methods, especially at high b‑values where SNR is low.

Computationally, BM4D‑PC adds the cost of a global PCA and multiple BM4D runs (one per PC). However, the multichannel BM4D implementation and reuse of block‑matching coordinates mitigate memory and runtime overhead, making the approach feasible on modern workstations. Because the pipeline is model‑based and does not require training data, it can be readily applied to a wide variety of acquisition protocols, coil configurations, and field strengths without retraining.

In summary, the contributions of this work are: (1) a comprehensive noise model for dMRI that includes spatial correlation and PSD; (2) a novel integration of global PCA with BM4D to exploit redundancy across the diffusion dimension; (3) automatic estimation of noise maps and PSD directly from the data; and (4) extensive validation demonstrating state‑of‑the‑art performance on both simulated and real datasets. Future directions include extending BM4D‑PC to multi‑TE or multi‑band acquisitions, coupling it with advanced microstructural modeling, and accelerating the algorithm via GPU implementation for near‑real‑time clinical use.