The image-quality transfer framework [1,2] transfers information from high quality to lower quality data sets. Preliminary work demonstrates its utility in two separate tasks. The first is super-resolution of diffusion tensor imaging (DTI), where [1,2] show major improvements over interpolation in predicting high-resolution images. The second is parameter mapping, where IQT predicts NODDI maps, which require multiple b-value shells, from DTIs fitted to single-shell data.

Here we demonstrate the benefits of IQT within the downstream application of tractography. We show that, through IQT, we can recover thin, nearby tracts from a data set acquired at relatively low (2.5mm isotropic) spatial resolution. Specifically, we focus on projections arising from the hand motor area in the cortex [3], that have been distinguished before using high resolution data from the Human Connectome Project (HCP) [4] (see figure 4 of [4], which shows that FSL tractography [5] can separate four pathways at 1.25mm resolution more readily than with 1.5mm or 2mm). We also show quantitatively that tractography on IQT super-resolved data via random-forest regression matches “ground truth” (tractography run at full resolution) better than global-linear IQT or interpolated data.

IQT super-resolution uses patch-regression to learn a mapping from patches, e.g. a 5x5x5 voxel neighbourhood, in low resolution images to the high-resolution patch, e.g. 2x2x2 neighbourhood, corresponding to the central voxel of the low-resolution patch (figure 1). The original work uses random-forest [6] regression of matched patch-pairs from downsampled high-resolution images.

Previous work [1,2] operates on patches of diffusion tensors, estimated from single b-value data. Here, to represent multi-shell data, we fit the Mean Apparent Propagator (MAP)-MRI [7] basis up to fourth order in each voxel and learn a mapping from high-resolution to low-resolution patches of MAP coefficients.

We use 8 diffusion MRI data sets from the HCP to train an IQT random-forest model. The specialist HCP scanner and imaging protocols provide data of uniquely high quality with 1.25mm isotropic resolution and 270 diffusion weighted images (DWIs) in three b-shells (1000, 2000, and 3000 $$$s/mm^2$$$) of 90 directions [4]. Input patches are 5x5x5 neighbourhoods of voxels downsampled by block averaging to 2.5mm resolution. Output patches are 2x2x2 neighbourhoods at 1.25mm resolution. Thus IQT learns a mapping that doubles the resolution in each dimension. The random forest contains 8 trees, each trained on ~750K patch pairs randomly sampled from the training images.

For testing, the input to IQT is a low-resolution 3D image of 22 MAP coefficients; the output is a high-resolution image of the same coefficients. Those coefficients predict the set of DWIs, which we input to the multi-fibre multi-shell probabilistic tractography algorithm in [5,8]. For comparison, we also generate high-resolution data sets by interpolation of each low-resolution DWI.

First, we test how well the learned IQT mapping works for 8 test subjects of the same HCP population (distinct from the training set). We seed tractography in the hand area of the motor cortex to obtain streamline probability maps at full 1.25mm resolution. For comparison, we downsample the original data to 2.5mm resolution and repeat the tractography on super-resolved 1.25mm data sets generated from IQT. Second, we assess the learned IQT mapping for datasets acquired on a different clinical scanner. We acquire two datasets from a Siemens Prisma 3T scanner, with 1.35mm and 2.5mm resolution and b-values and gradient directions matching the HCP protocol. We use IQT to super-resolve the 2.5mm data to 1.25mm and compare tractography on the original 2.5mm, 1.35mm, and IQT super-resolved 1.25mm data sets.

Figure 2 demonstrates that super-resolution via IQT with random-forest regression enables tractography to identify four separate pathways from the hand area to the thalamus, brainstem, spinal cord and putamen, whereas standard interpolation does not. Quantitative results (figure 3) show that the random-forest IQT leads to tractography results that best match those from full-resolution data consistently over 8 test subjects.

Figure 4 shows recovery of the four pathways in the Prisma data after IQT super-resolution to 1.25mm. These pathways are not all present in the raw 2.5mm data. The tractography maps from the IQT data set closely match those from the original 1.35mm data.

We show how to use IQT for super-resolution of diffusion MRI data sets that enable multi-shell, multi-fibre tractography. The method extends previous work by using the higher-order MAP signal-representation in place of the diffusion tensor model. Results show that IQT super-resolution benefits tractography more than standard interpolation. This demonstrates the potential for IQT to provide for standard data sets the specificity of tractography previously only available from highly specialized data.