MBD: Multi b-value Denoising of Diffusion Magnetic Resonance Images

TL;DR

A convolutional neural network approach is introduced that is called multi-b-value-based denoising (MBD), which allows denoising of diffusion images with high noise variance while avoiding blurring, and using just a small number input images.

Abstract

We propose a novel approach to denoising diffusion magnetic resonance images (dMRI) using convolutional neural networks, that exploits the benefits of data acquired at multiple b-values to offset the need for many redundant observations. Denoising is especially relevant in dMRI since noise can have a deleterious impact on both quantification accuracy and image preprocessing. The most successful methods proposed to date, like Marchenko-Pastur Principal Component Analysis (MPPCA) denoising, are tailored to diffusion-weighting repeated for many encoding directions. They exploit high redundancy of the dataset that oversamples the diffusion-encoding direction space, since many directions have collinear components. However, there are many dMRI techniques that do not entail a large number of encoding directions or repetitions, and are therefore less suited to this approach. For example, clinical dMRI exams may include as few as three encoding directions, with low or negligible data redundancy across directions. Moreover, promising new dMRI approaches, like spherical b-tensor encoding (STE), benefit from high b-values while sensitizing the signal to diffusion along all directions in just a single shot. We introduce a convolutional neural network approach that we call multi-b-value-based denoising (MBD). MBD exploits the similarity in diffusion-weighted images (DWI) across different b-values but along the same diffusion encoding direction. It allows denoising of diffusion images with high noise variance while avoiding blurring, and using just a small number input images.

Figures (14)

• Figure 1: Configuration and architecture of the MBD, CNNe and N2N convolutional neural networks. Note the differences in the input, while the architecture and target data are the same. $C$ is the number of input images (channels) for CNNe. $C=2$ case is presented.
• Figure 2: Example axial slice from the clinical brain dMRI, involving b-values of 0, 300 and 1000 s/mm$^2$.
• Figure 3: Example axial slice from the STE brain dMRI, involving b-values of 0, 1000 and 4000 s/mm$^2$.
• Figure 4: Example axial slice from synthetic brain DTI. Noiseless reference (top) and noisy images with $\sigma= 7\%$ of the true white matter intensity at b=0 (bottom) are shown. Several differently shaped lesions are visible, with $f=0.75$, $D_1=400\frac{mm^2}{s}$, $D_2=600 \frac{mm^2}{s}$, $\Delta T_2=15$ ms. They are hyperintense at all b-values compared to normal white matter. Exponential SNR drop is well visible for the b4000 image.
• Figure 5: Comparison of MBD, N2N, CNNe, MPPCA and ALGe denoising of b1000 clinical dMRI images. Rows, from top: NEX = 1 images, NEX = 32 images, NEX=32 images (zoomed view), NEX = 32 absolute residuals. MBD shows accurate denoising and edge-preservation, even for NEX=1 images.