Introduction

Dynamic contrast-enhanced MRI (DCE-MRI), which is used to obtain the information about permeability of vascular, hemodynamic and components’ proportion of tissue,¹ is widely applied in clinical research and diagnosis of tumors, neurodegeneration disease.²

There are many kinetic models proposed to analyze the DCE-MRI data, however, the curve fitting of the DCE‐MRI time‐courses is error-prone due to the non-white, spatially-dependent noise.³ Hence, for DCE-MRI, many approaches have been put forward to reduce the noise, such as the gaussian low pass filter (GLPF), the dynamic nonlocal mean (DNLM),⁴ the bilateral filter.⁵ What’s more, due to the similar characteristic, the 2D kinetics-induced bilateral filter (KIBF-2D) and the nonlocal mean based on spatiotemporal patches (NLM-ST),^6,7 which are used in denoising of dynamic PET images, are also valuable for DCE-MRI.

In this study, we aim to compare the different denoising approaches for DCE-MRI, which include GLPF, NLM-ST,⁶ DNLM,⁴ KIBF-2D,⁷ and KIBF-3D modified from KIBF-2D, and in the performance of denoising DCE-MRI data.

Methods

The conventional nonlocal mean (NLM) re-establish the given pixel signal by using the similarity, which is defined as the intensity difference, in the neighborhoods.⁸ NLM-ST is modified for DCE-MRI, which focuses on spatial information.⁶ On the other hand, DNLM moved the focus of NLM from spatial domain to time domain to denoise DCE-MRI data.⁴ Unlike the GLPF whose weighted matrix derived from distance difference and the NLM whose weighted matrix derived from intensity difference, the KIBF, which is based on bilateral filter, combines these two characteristics.⁷ To increase the utilization of spatial information, we extended KIBF’s neighbor windows from 2D to 3D, which is called KIBF-3D. Please refer to the references for specific formulas due to words limitations.^4,6,7

Ten high grade glioma (HGG) patients and ten solitary brain metastasis (SBM) patients were scanned on a 3.0T MRI instrument (Magnetom Skyra, Siemens Healthcare, Erlangen, Germany). DCE-MRI data were acquired with 3D CAIPIRINHA-Dixon-TWIST sequence with the parameters: FOV, 340 × 340 × 120 mm³; resolution, 0.8×0.8×1.5 mm³; flip angle, 10°; TR, 6 msec; TE, 1.3 msec; temporal resolution, 4.5 seconds; 120 frames (~ 9 min).

Simulations were performed to optimize the setting of NLM-ST, DNLM, 2D and 3D KIBF. Specifically, an artificial DCE-MRI brain phantom with DCE-MRI signal generated from the extended Tofts (eTofts) model with Rician noise added.⁹ The brain was divided into 121 regions base on MNI152 T1-weighted template images and the K^trans, v_e, v_p and quantitative T1 of each brain region were randomly picked from reasonable ranges.¹⁰ At last, the noisy DCE-MRI signal were denoised with different approaches and then fitted with the eTofts model. The bias, dispersion and peak signal-to-noise ratio (PSNR) of the eTofts parameters were then used to evaluate the denoising approaches.

Those denoising methods were also applied onto the experimental DCE-MRI data and the noisy and denoised DCE-MRI data were then fitted by eTofts model. As there is no ground truth, the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR, tumor vs normal appearing white matter (NAWM)) were used in the assessment. The SNR and CNR were defined as follows,
$$S N R = \frac { mean [ \textbf{I} _{r e s u l t } ( t u m o r ) ] } { s t d [ \textbf{I}_ {r e s u l t } ( NAWM ) ] }$$
$$C N R = \frac { mean [ \textbf{I} _{r e s u l t } ( t u m o r ) ]-mean [ \textbf{I} _{r e s u l t } ( NAWM) ] } { s t d [ \textbf{I}_ {r e s u l t } ( NAWM ) ] }$$
where $$$\textbf{I} _{r e s u l t }$$$ is the K^trans or v_p of the tumor ROIs or NAMW ROIs, which were manually selected by two experienced neuroradiologists together.

Results and Discussions

Parametric maps from the simulated brain DCE-MRI phantom preprocessed with different denoising approaches are shown in Figure 1A. Compared with other algorithms, the KIBF-2D and KIBF-3D reduce the noise significantly and reserve subtle information best, through bias also exists in some region. In Figure 1B&C, an typical example of the DCE-MRI time-course signal pre- and post-processed by the denoising methods are shown, in which KIBF-3D approach produces the results closest to the ground truth. Further quantitative evaluation also confirms that KIBF-3D achieves the smallest bias and dispersion, and has the greatest PSNR (Figure 2).

The eTofts parametric maps from one HGG subject’s DCE-MRI data are shown in Figure 3A. In visual inspection, KIBF-3D performs best, especially in K^trans map in which the K^trans values in normal appearing brain area are close to minimum and more uniform. In addition, typical examples of pre- and post-denoised DCE-MRI time-course signal of one voxel are shown in Figure 3B&C, in which the difference between pre- and post-denoised signal with KIBF shows closest to white noise. Further quantitative evaluation also shows that KIBF-3D outperforms other methods in the SNR and CNR of of K^trans and v_p, though NLM-ST shows similar performance on v_p in SBM patients’ data (Figure 4).

Conclusion

To achieve denoised DCE-MRI data with fine structural details, we compare different denoised methods. Both simulation and in vivo HGG and SBM study have shown the KIBF-3D algorithm can achieve significant advantage compared with other algorithms in terms of quantitative evaluation and visual inspection.

Acknowledgements

No acknowledgement found.