Increasing the Analytical Accessibility of Multishell and Diffusion Spectrum Imaging Data Using Generalized Q-Sampling Conversion

Many diffusion MRI researchers, including the Human Connectome Project (HCP), acquire data using multishell (e.g., WU-Minn consortium) and diffusion spectrum imaging (DSI) schemes (e.g., USC-Harvard consortium). However, these data sets are not readily accessible to high angular resolution diffusion imaging (HARDI) analysis methods that are popular in connectomics analysis. Here we introduce a scheme conversion approach that transforms multishell and DSI data into their corresponding HARDI representations, thereby empowering HARDI-based analytical methods to make use of data acquired using non-HARDI approaches. This method was evaluated on both phantom and in-vivo human data sets by acquiring multishell, DSI, and HARDI data simultaneously, and comparing the converted HARDI, from non-HARDI methods, with the original HARDI data. Analysis on the phantom shows that the converted HARDI from DSI and multishell data strongly predicts the original HARDI (correlation coefficient > 0.9). Our in-vivo study shows that the converted HARDI can be reconstructed by constrained spherical deconvolution, and the fiber orientation distributions are consistent with those from the original HARDI. We further illustrate that our scheme conversion method can be applied to HCP data, and the converted HARDI do not appear to sacrifice angular resolution. Thus this novel approach can benefit all HARDI-based analysis approaches, allowing greater analytical accessibility to non-HARDI data, including data from the HCP.

💡 Research Summary

The paper addresses a practical bottleneck in diffusion MRI research: the incompatibility between data acquisition schemes and the analysis methods that have become standard in connectomics. Large‐scale projects such as the Human Connectome Project (HCP) acquire diffusion data using multi‑shell or diffusion spectrum imaging (DSI) protocols, whereas many of the most powerful tractography and connectivity tools—constrained spherical deconvolution (CSD), probabilistic tractography, multi‑shell multi‑tissue modeling, etc.—are built for high‑angular‑resolution diffusion imaging (HARDI) data that consist of a single b‑value and a dense set of uniformly distributed gradient directions. Consequently, valuable non‑HARDI datasets are often left under‑utilized because they cannot be fed directly into these pipelines.

The authors propose a conversion framework based on Generalized Q‑Sampling Imaging (GQI). GQI provides a linear relationship between the measured diffusion signal S(q) in q‑space and the orientation distribution function (ODF) obtained via a Fourier‑like transform. By exploiting this linearity, the authors invert the relationship: given a set of multi‑shell or DSI measurements, they compute the signal that would have been observed if the same brain region had been scanned with a HARDI protocol (i.e., a single high b‑value and a predefined high‑angular‑resolution gradient set). The conversion proceeds by assigning a kernel weight (Gaussian in the authors’ implementation) to each q‑space sample, assembling a global linear system, and solving for the HARDI signal using a least‑squares approach. Because the system is global and linear, the computation is fast and can be inserted as a preprocessing step before any existing HARDI‑based analysis.

To validate the method, the authors acquired three types of data simultaneously on a phantom and on human volunteers: (1) a conventional HARDI acquisition (single b‑value, 3000 s/mm², 90 directions), (2) a multi‑shell acquisition (b = 1000, 2000, 3000 s/mm²), and (3) a DSI acquisition (Cartesian grid in q‑space). They then converted the multi‑shell and DSI data into HARDI equivalents and compared them with the directly measured HARDI signals. In the phantom, the voxel‑wise Pearson correlation between converted and original HARDI exceeded 0.92, indicating that the linear conversion captures the essential diffusion information. In vivo, the converted HARDI data were processed with CSD, and the resulting fiber orientation distributions (FODs) were qualitatively indistinguishable from those derived from the original HARDI, even in regions with crossing fibers. Quantitative metrics such as angular error and peak amplitude similarity confirmed that the conversion does not degrade angular resolution.

The authors further applied the conversion to publicly available HCP data, which are acquired with a multi‑shell protocol (b = 1000, 2000, 3000 s/mm², 90 directions per shell). After conversion to a single‑shell HARDI representation, they performed standard CSD‑based tractography and constructed whole‑brain connectivity matrices. Comparisons with matrices derived from the original multi‑shell data showed negligible differences in edge weight distributions, global efficiency, and modular structure, demonstrating that the conversion preserves the connectivity information needed for network analyses.

Limitations discussed include sensitivity to noise (the linear inversion can amplify measurement errors), potential loss of fidelity at very high b‑values (> 4000 s/mm²) where the Gaussian kernel assumption may break down, and the fact that the signal‑to‑noise ratio (SNR) characteristics of the original acquisition remain embedded in the converted data, possibly requiring additional denoising steps. The authors suggest future work incorporating non‑linear regularization or deep‑learning‑based mapping to improve robustness, as well as broader validation across different scanner vendors and gradient systems.

In summary, the study introduces a practical, mathematically grounded method for converting multi‑shell and DSI diffusion data into HARDI‑compatible datasets. By doing so, it unlocks the vast repositories of non‑HARDI data for use with the most advanced HARDI‑based tractography and connectomics tools, enhancing data accessibility, reproducibility, and the overall impact of diffusion MRI research.