Interpolation of DWI prior to DTI reconstruction, and its validation

Introduction: The resolution of diffusion weighted imaging (DWI) is mainly limited by hardware constraints and a negative relationship between image resolution and signal-to-noise ratio (SNR). Moreover, SNR can only to a limited extent be improved by repeating the measurements. Super-resolution (SR) is a technique that fuses information from a series of interpolated low resolution images to create a single high resolution image. A prerequisite is that the multiple low resolution images contain statistically independent data of the same object, e.g. capture a static object displayed from different spatial viewpoints. Since this is the case for HARDI DWI, we hypothesized that SR can extract anatomical details from HARDI DWI data that are obscured by partial volume effect (PVE) and thus, are currently only visible when DWI is acquired at a higher spatial resolution, from such data sets. In contrast to previous SR methods which are limited to special acquisition sequences or require the introduction of additional complex SR reconstruction methods applied after fibre reconstruction, we introduce a simple and general SR framework for HARDI DWI. We propose that by incorporating standard interpolation of the acquired (low resolution) DWI data, and then applying the diffusion tensor model for reconstruction (though any fibre reconstruction method can be used), we are able to significantly increase the spatial resolution of the final image and hereby to visualize more anatomical details. Preliminary results from SR interpolation of low-resolution high-quality ex vivo HARDI DWI datasets acquired on perfusion fixed monkey brain are validated against high resolution acquisitions of the same datasets. Method: Ex vivo imaging of an 82 month healthy perfusion fixed Vervet monkey brain was performed on an experimental 4.7 tesla Varian Inova scanner. Two whole brain DWI datasets were acquired with isotropic voxels of size 0.8 and 0.5 mm respectively (no gap). As a validation golden standard a high-resolution DWI dataset was also acquired, comprising 30 coronal slices partly covering the brain, with in-plane resolution 0.25 x 0.25 mm, slice thickness 0.5 mm and 0.5 mm gap. High SNR was ensured by NEX of 1, 4 and 7, resulting in a final SNR of 52, 48 and 16 respectively. For all datasets: TR=5500 ms, TE=47 ms, 3 x b0 volumes and 61 dw directions. For the ex vivo DWI, the setup of Dyrby et al. (2010) was employed, and from there a b-value of ~4300 s/mm was selected. Conditioned airflow surrounded the brain during scanning. All data sets were acquired in a single scanning session and no additional processing was needed. Ethical rules concerning care and handling of live animals were followed. Analysis: The suggested SR framework was realised by interpolating the original DWI data followed by fibre reconstruction using the diffusion tensor model. We define the SR factor as the ratio between the volume of original and the interpolated voxels, hence unit SR factor is conventional DTI. Data was resliced using 7-order B-spline interpolation using SPM8, and DTI matrices were calculated using Camino. Only interpolation to isotropic voxels was employed which ensured optimal rotational invariance in the SR-DWI data. Results: DTI reconstruction from high resolution DWI data showed anatomical details not visible at conventional resolutions (Fig 1, b). Using our simple SR approach upon the low resolution DWI datasets, similar anatomical details became more apparent (Fig 1, c & d vs. e & f). An SR of 27 upon 0.8 mm voxels (Fig 1, c & e) extracts many of the hidden anatomical details seen in (Fig 1, b), though with slight smoothing when compared with SR=8 upon the 0.5 mm data (Fig 1, f). The latter is similar to the golden standard (Fig 1, b).