Super-resolution (SR) is a concept using several existing low-resolution (LR) images to construct high-resolution (HR) images through signal processing of the existing captured images. The sparsity-guided SR is a robust approach recently used in enhancing the regular picture resolution¹ based on the idea that the sparse signal representations can be correctly recovered from the down-sampled image patch to generate a high resolution image patch. This method has been proved to be superior to other SR methods and more robust to image noise. In this work, Sparsity-guided SR Imaging will be adapted to reconstruct clinical MPRAGE in order to improve the tissue segmentation of the brain.

Image Acquisition: Brain tissue segmentation was based a set of standard MPRAGE images (TR/TE=2300/2.98ms, FA=9°, voxel size=1x1x1mm³, 1 average) acquired at 3T MR, which were considered as low resolution (LR) images. In order to compare the efficacy of the SR algorithm, an additional set of high-resolution (HR) MPRAGE images (TR/TE: 2300/3.59ms, FA=9°, voxel size = 0.5x0.5x1 mm³) were acquired with half the pixel size and were used as target resolution. To compensate SNR, 2 averages were applied on HR MPRAGE imaging with acquisition time already 4 times longer than the standard MPRAGE imaging although the SNR was not fully recovered.

SR Image processing: The proposed algorithm, initially proposed by^1-3, exploits the similarity of the sparse representation (α) of the same object with two different resolution. This is, DH XH=DL XL=α, where DH and DL are the HR and LR representation bases (dictionaries) respectively and XH and XL are the HR and LR images respectively. The training of the dictionaries part was previously reported^1-3. To recover the HR image from the LR image, the method estimates the relationship between DL and DH and applies to LR patches for recovering HR patches.

Segmentation: Images were segmented into gray matter (GM), white matter (WM) and cerebrospinal fluid (CSF) using the voxel-based morphometry toolbox version 8 (VBM8; http://dbm.neuro.uni-jena.de/vbm/). Whole brain volumes of GM and WM as well as region-of-interest (ROI) volumes were measured and compared between 3 sets of images (a) standard LR, (b) target HR, and (c) reconstructed SR.

As shown in Figure 1, SR images showed much improved delineation of all structures with sharper edges and increased probability assigned for each tissue type on the segmented images. Three regions of interest (i.e., cerebellum, cortical and deep gray matter) were measured and compared of their volumes and relative changes on representative LR, HR, and SR images (Table 1 and Figure 2). There is much improved separation of highly folded cerebellar gyri from WM (i.e., arbor vitae) on SR than original LR images with average of 15% decrease of volume on SR due to decrease of partial volume with WM, and such changes are close to changes on targeted HR images (Figure 3). Cortical GM showed 2% decrease of GM on SR compared to LR (Figure 4). However, this was not consistent across subjects. Segmented images based on SR also showed improved deep GM delineation of intranuclear WM bundles (Figure 5) compared to both LR and HR images, which suggests that HR images may have suffered compromised SNR when higher resolution was applied.The SR method proposed here aims to reach a higher resolution for improved brain tissue segmentation and to overcome the inherent MRI image resolution limitations due to longer acquisition time, reduced SNR, and potential motion artifacts to acquire higher resolution imaging. The segmentation results suggest that SR method can largely increase the probability of small structures and edge definition compared to original LR, therefore to increase accuracy of tissue segmentation. The results on SR and HR are comparable and in some brain regions SR is even better than HR because unlike SR, HR can suffer from compromised SNR. Whole-brain GM volumes of SR images were closer to HR than LR; this may suggest that the SR algorithm may improve whole-brain GM segmentation.