We evaluate the generalisability of image-quality-transfer (IQT) for diffusion tensor imaging (DTI) super-resolution. Specifically, we show that the reconstruction quality of high-resolution in-vivo human diffusion tensor imaging (DTI) maps from the Human Connectome Project (HCP)^3,9 degrades only slightly as we move through IQT models trained on i) other HCP data from similar subjects; ii) HCP data from a distinct demographic; iii) data from a distinct acquisition protocol and very different demographic (HCP Lifespan dataset⁷); and, remarkably, iv) data from ex-vivo fixed monkey brains⁴.

The IQT^1,2 framework embeds machine learning in image reconstruction with the aim of incorporating information from bespoke high-quality datasets into images derived from standard acquisitions. For example, IQT super-resolution learns a mapping from low-resolution to high-resolution images from matched pairs obtained by downsampling high-resolution images. The approach strongly outperforms interpolation in recovering high resolution DTIs from low resolution data^1,2. Previous work^1,2 also shows prediction of NODDI⁸ maps, which normally require two b-value shells, from single-shell b=1000s/mm² DTI data. The framework has broad potential in a variety of other applications. However, one key question, which we address here, is the ability of IQT to generalise to unfamiliar situations (e.g. in clinical data) that are not directly represented in the training data.

IQT super-resolution^1,2 uses random-forest⁵ 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 training phase takes a set of high resolution images, downsamples to approximate a low-resolution acquisition, draws a sample of matched high and low-resolution patches, and constructs a mapping by non-linear regression. That mapping then operates on low-resolution images to reconstruct a corresponding high-resolution image patch by patch.

To demonstrate generalisability of IQT, we test its ability to reconstruct high-resolution DTIs after training on various different kinds of data. First we use data from numerous subgroups of the HCP cohort. Subgroups 2 (HCP2) and 3 (HCP3) are homogeneous but maximally different: HCP2 subjects are white, right-handed, females in their 20s; HCP3 subjects are black, left-handed, males in their 30s. HCP1 is heterogeneous containing subjects with all combinations of those traits. Support vector machine classification⁶ using brain images from HCP2 and HCP3 strongly discriminates the two groups. We use distinct training and test sets, each of 8 subjects, from both HCP1 and HCP2 subgroups, as well as a test set only from HCP3, also containing 8 subjects. Training and testing uses DTI maps estimated from the 90-direction b=1000s/mm² shell.

We also use training sets from i) the HCP Lifespan dataset and ii) a set of fixed monkey brain images. The Lifespan data comes from 26 subjects with an age range of 4-75 years; the imaging protocol has a b=1000s/mm² shell similar to the HCP data but with 1.5mm rather than 1.25mm isotropic voxels. The Monkey data has images from 13 perfusion-fixed vervet monkey brains acquired with an ex-vivo imaging protocol⁴. We use a single b=3151s/mm² shell with 87 directions.

To test, we reconstruct each image in the HCP1, HCP2, and HCP3 test sets after downsampling by block averaging. We compute the root-mean-squared DT error (DT-RMSE) between the IQT-reconstructed DTI map and the ground-truth from fitting to the original full-resolution acquisitions.

Figure 2 demonstrates generalisability of IQT within the HCP cohort. Reconstruction errors vary among test sets, but, importantly, depend little on the training set.

Figure 3 shows HCP1 errors scores after training on the HCP, Lifespan, and Monkey data. Although errors from the Lifespan and Monkey sets are significantly higher than from HCP training, both still produce good approximations to the ground truth. Figure 4 compares IQT-reconstructed DTI maps from different training sets. To emphasise: the high resolution DTI map in the top right of figure 4 is effectively a patchwork of dead monkey brain image-pieces. Remarkably, it is hardly distinguishable from the reconstruction using HCP training data.

We demonstrate generalisability of IQT super-resolution of DTI by training and testing on datasets that vary in race, age, gender and handedness (HCP); acquisition protocol, resolution and age (Lifespan); and protocol, species and tissue state (Monkey). The experiments show promising robustness of the IQT framework to training set, which demonstrates the potential to exploit the technique even in situations when data comes from unknown or unfamiliar subjects or samples. Future work still needs to assess robustness in the presence of pathology.