High-Resolution Susceptibility Weighted Images Derived from fMRI Images using Super-Resolution Reconstruction
Weiran Deng1, Michael Herbst1, and V. Andrew Stenger1

1University of Hawaii JABSOM, Honolulu, HI, United States


A subset of fMRI images is used to reconstruct a structural image at a higher resolution using a Super-Resolution (SR) reconstruction method. The subset of fMRI images are selected such that the translation and rotation between the shots are less than the pixel size and therefore useful for modeling the pixel characteristics. The preliminary results demonstrate the feasibility of reconstructing a structural image with susceptibility contrast from a subset of fMRI images.


To reconstruct a High-Resolution Susceptibility Weighted Image (SWI) from a subset of low-resolution functional MRI (fMRI) data using Super-Resolution (SR) reconstruction.


Subject head motion is inevitable during fMRI scans. Even when the subject makes an effort to remain still, sub-voxel involuntary motions are still impossible to control. However, such motions are useful for reconstructing a high-resolution image using SR reconstruction from the multiple low resolution fMRI images. There are existing studies that use SR for DTI white matter fiber track visualization1 and to enhance imaging resolution in general2. This abstract presents preliminary results of a high-resolution SWI reconstructed from a subset of fMRI data. The high-resolution image can then be used as a template over which the activation from the low resolution scans can be overlaid.


Conceptually, a high-resolution image can be reconstructed form a series of low-resolution images provided these images have sub-pixel shifts with respect to each other. The spatial characteristics of the voxel can be modelled and used to reconstruct an image with a higher resolution. The observation model of the i-th low-resolution fMRI image yi, is the result of applying a series of linear operators (Di, Fi, Hi) and the addition of a noise term ni: $$y_{i}=D_{i}F_{i}H_{i}z+n_{i}$$, where Di, Fi, and Hi represent the down-sampling, geometrical transformation, and blurring, respectively. The reconstruction of the high-resolution image z is a classical linear inverse problem, which can be solved using methods such as non-uniform interpolation, the frequency domain approach, projection-onto-convex sets, or regularized methods3.


A resting state fMRI scan was performed on a 3T Siemens TimTrio scanner using a 32-channel head coil. The sequence was a Simultaneous Multi-Slice EPI (TE/TR=30/1000ms, FOV=220mm, 64 2mm slices, FA=55o). Motion parameters (translation and rotation) were estimate using FSL’s FLIRT program4. Sixteen volumes were selected from the fMRI data and the SR reconstruction was permed using a MATLAB SR program5. A structure-adaptive normalized convolution approach was used for the SR reconstruction at 2mm resolution6. The resting state Independent Component Analysis (ICA) analysis was performed using FSL’s MELODIC program7.


Figure 1 shows image translation (top) and rotation during the fMRI scan. The black dashed box annotates the time points of the 16 volumes selected for the SR reconstruction. Note that the translation (less than 2mm pixel size) and rotation were small enough to be useful for SR. Figure 2 shows a comparison of a LR image (a), an HR image (b) interpolated from the LR image using a bilinear interpolation method, and an SR image (c), respectively. A resting state activation map from the default mode was overlaid on top of the bilinear interpolated image and the SR reconstructed image and shown in Figures 3a and 3b, respectively.


There are several ways to improve the SR reconstruction presented here. The SR images were reconstructed slice-by-slice from the LR fMRI data for computational ease. Better results may be obtained if the SR reconstruction is performed on the whole 3D volume using 3D translation and rotation parameters. Second, the alignment from motion was estimated using FLIRT. Because the SR reconstruction highly depends on the accuracy of the motion parameters, the reconstruction can be improved using parameters measured accurately using a MRI-compatible motion-tracking system8. Third, it is unclear if the variation of the voxels from BOLD hemodynamics contributes to the quality of the SR reconstruction. A potential solution is to use a remove hemodynamic trend before the SR reconstruction.


The preliminary results presented here demonstrate the feasibility of deriving a high-resolution image using a subset of fMRI images. Future work will investigate options to improve the image quality of SR reconstruction.


Work supported by the NIH grants R01DA019912, R01EB011517, and K02DA020569.


1. Peled et al, Superresolution in MRI: application to human white matter fiber tract visualization by Diffusion Tensor Imaging, Magn. Reson .Med. 45:29-35, 2001.

2. Plenge et al, Super-Resolution methods in MRI: Can they improve the trade-off between resolution, signal-to-noise ratio, and acquisition time? Magn. Reson. Med. 68:1983-1993, 2012.

3. Park et al, Super-Resolution image reconstruction: a technical overview, IEEE Signal Processing Magazine, 20(3):21-36, 2003.

4. Jenkinson et al, A global optimization method for robust affine registration of brain images. Medical Image Analysis, 5(2):143-156, 2001.

5. Vandewall et al, A frequency domain approach to registration of aliased images with application to super-resolution. EURASIP Journal on Applied Signal Processing, vol. 2006:1-14, 2006.

6. Pham et al, Robust fusion of irregularly sampled data using adaptive normalized convolution, EURASIP Journal on Applied Signal Processing, vol. 2006:1-12, 2006.

7. Beckmann et al, Probabilistic independent component analysis for functional Magnetic Resonance Imaging, FMRIB Technical Report TR02CB1.

8. MacLaren et al, Measurement and correction of microscopic head motion during Magnetic Resonance Imaging of the brain. PLOS One, 7(11):e48088.


Figure 1. The top and bottom plots show translation and rotation during the fMRI scan. The segment marked by the dashed black box is the subset of fMRI images selected for SR reconstruction.

Figure 2. (a), (b), (c) show a low resolution fMRI image (2mm), its bi-linearly interpolated high-resolution image (1mm), and the SR image (1mm), respectively.

Figure 3. (a) and (b) show the activation corresponding to the default mode overlaid on top of the bilinearly interpolated high-resolution image and the SR image.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)