5477
High Resolution fMRI Data De-noising Technique Using Spatio-Temporal Diffusion Filter1Department of Biomedical Engineering, Amirkabir University of Technology, Tehran, Iran (Islamic Republic of), 2School of Cognitive Sciences, Institute for research in fundamental sciences (IPM), Tehran, Iran (Islamic Republic of)
Synopsis
Diffusion-based filtering approaches are widely used in MR image de-noising literature but only a few studies, utilize this technique for fMRI applications. In these studies, 1D & 2D diffusion filters were applied using temporal and spatial information separately. Here, a novel
Introduction
Anisotropic filtering is the effective method to reduce noise and increase functional signal in fMRI studies. Researchers have proposed different methods for adaptive smoothing, such as maximum energy ratio statistic [1], model-based spatial filters [2], nonlinear diffusion based filters [3], etc. Diffusion-based filtering approaches are widely used in MR image de-noising literature but only a few studies, utilize this technique for fMRI applications [3-5]. In these researches, one and two dimensional (1D & 2D) diffusion filters were applied using temporal and spatial information separately [4]. Since routine low spatial-resolution fMRI data generally do not have sufficient contrast between gray matter and other tissues, spatial information could not be informative unless we utilize the structural images co-registered to fMRI data [6]. However, for higher spatial resolution imaging techniques which has the high image SNR like SSFP technique, it would be more effective to consider both temporal and spatial information simultaneously to smooth data with no need to use structural data. Here, we propose a 3D diffusion filtering technique to boost functional sensitivity along with persevering the boundaries of the active regions in high-resolution SSFP data utilizing both spatial and temporal information of neighboring voxels. The results on the experimental SSFP data shows the ability of the proposed technique in improving functional sensitivity as well as preserving the edges of active regions.Material & Methods
The mathematical model of diffusion filtering technique was initially proposed in [7] inspired by a diffusion process. In this model, a local smoothing function (C) defined as the diffusion tensor; the following equation used to smooth three dimension data (I) along the diffusion function (C):
$$\frac{\partial }{\partial t}I(x,y,t)=div(C(x,y,t))\triangledown I(x,y,t)) [1]$$
Where C is denoted as a 3D function which diffuses into both temporal and spatial directions. Here, a decreasing function was selected for it as follows:
$$C(x,y,t)=\frac{1}{(1+abs(\frac{\triangledown I(x,y,t)}{K})^2)} [2]$$
Parameter K is the gradient modulus that controls the edge preserving condition which was experimentally selected as 500. To evaluate the proposed method, it was compared with two other kernel-based de-noising techniques. The first one uses the standard Gaussian kernel with a window size of 3*3, zero mean and standard deviation of 1. The second one applies the temporal diffusion kernel proposed in [4]. For the temporal and the proposed diffusion methods, the same K was used with a window size of 3*3 and with the iteration of 10. For SSFP data, prior to proposed technique, preprocessing steps (motion correction, brain extraction, and a high-pass filter) were performed in FSL (Figure 1).
High-resolution SSFP-fMRI data was acquired based on CE-FAST sequence [8] using 7T Siemens scanner which has high SNR images without banding artifact, and spatial specificity close to Spin-Echo. Five randomly healthy subjects were chosen with the imaging parameters as follows: in-plane resolution=1.25x1.25mm2, thickness=5mm, TE/TR=10.9/14.1ms, FA=23°, volume-TR=8s, total duration=320s. In all fMRI acquisitions, a combined visual-motor task was performed simultaneously.
Results
First, data was pre-processed by pipeline represented in Figure 1. The mentioned smoothing techniques were then applied to preprocessed data and the correlation coefficients of time-courses with the stimulus pattern were calculated. To compare the level of functional sensitivity, the mean value of correlation coefficients over 1000 voxels with the highest coefficients were calculated for all techniques which presented in Table 1. Furthermore, the maximum correlation coefficient was reported that shows our filtering technique was successfully prevented the signal leakage from the highly activated voxels. The correlation maps, overlaid on functional images, of one selected slice of the representative subject are indicated in Figure 2 for different methods. As shown in this figure, while preserving the boundaries of the active parts, our method increases the level of activation in the related brain regions.Disscusion
The anisotropic filtering is a more effective approach to smooth fMRI data compared to isotropic one [2]. In anisotropic approach, spatial information can increase the performance of the de-noising technique. However, fMRI image mainly has a poor spatial resolution which practically makes the use of spatial information ineffective [9]. Here, we propose the 3D diffusion kernel to improve both functional sensitivity and preserve the functional resolution in SSFP-fMRI technique. According to Table 1, the proposed technique has the average values of maximum and mean correlation coefficients equal to 0.934 and 0.735 respectively over five subjects which is the highest in comparison to Gaussian and temporal diffusion kernels, indicating its better performance.Conclusion
To conclude, spatio-temporal diffusion methods, that locally de-noise fMRI data are more effective techniques for de-noising high spatial-resolution fMRI data which benefit from increasing the sensitivity, avoiding remove small regions and preserving functional resolution.Acknowledgements
The authors thank Prof. David G. Norris for facilitating the data acquisition. This work was supported by Iranian Council for Cognitive Sciences and Technologies (COGC) by grant no. 3198.References
[1] Hossein-Zadeh, G-A., Babak A. Ardekani, and Hamid Soltanian-Zadeh. "Activation detection in fMRI using a maximum energy ratio statistic obtained by adaptive spatial filtering." IEEE transactions on medical imaging 22.7 (2003): 795-805.
[2] Monir, Syed Muhammad Ghazanfar, and Mohammed Yakoob Siyal. "Denoising functional magnetic resonance imaging time-series using anisotropic spatial averaging." Biomedical Signal Processing and Control 4.1 (2009): 16-25.
[3] Kim, Hae Yong, Javier Giacomantone, and Zang Hee Cho. "Robust anisotropic diffusion to produce enhanced statistical parametric map from noisy fMRI." Computer Vision and Image Understanding 99.3 (2005): 435-452.
[4] Amir, Muhammad, Jawad A. Shah, and Suheel A. Malik. "De-noising of Functional Magnetic Resonance Imaging (fMRI) data using Nonlinear Anisotropic 1D and 2D filters." International Journal of Advanced Research in Computer Science 4.4 (2013).
[5] Khaliq, Amir A., et al. "De-Noising Functional Magnetic Resonance Imaging (fMRI) Data Using an Exponential Gradient Filter." vol 18 (2013): 1349-1356.
[6] Nam, Haewon, et al. "A method for anisotropic spatial smoothing of functional magnetic resonance images using distance transformation of a structural image." Physics in medicine and biology 56.15 (2011): 5063.
[7] Perona, Pietro, and Jitendra Malik. "Scale-space and edge detection using anisotropic diffusion." IEEE Transactions on pattern analysis and machine intelligence 12.7 (1990): 629-639.
[8] Malekian, Vahid, et al. “A Robust SSFP Technique for fMRI at Ultra-High Field Strengths”, Proc. Intl Soc Mag Reson Med. Vol. 25. 2017.
[9] Monir, Syed Muhammad, and Mohammed Yakoob Siyal. "Iterative adaptive spatial filtering for noise‐suppression in functional magnetic resonance imaging time‐series." International Journal of Imaging Systems and Technology 21.3 (2011): 260-270.
Figures
