What new or emerging applications of diffusion MRI are you most excited about in 2026, and why?

I am most excited about the application of diffusion MRI in grey matter. Alterations in the structure of grey matter are implicated but still not well understood in aging and almost all brain pathologies. Diffusion MRI has been used in thousands of studies to characterize white matter microstructure. Modern dMRI techniques are now beginning to allow us to characterize the more complicated structure of grey matter. Understanding the microstructural changes that underlie deficits in function, such as changes soma or fiber populations, and how they evolve over time, stands to greatly improve our understanding of neurological aging and disease.

What recent advances, whether in gradient hardware, diffusion encoding, or reconstruction, have made these applications more feasible? How are these new applications influencing the way we think about acquisition design?

There are many! Gradient hardware advances are certainly expanding the available design-space for diffusion encoding, enabling advanced techniques within a reasonable echo-time. In my lab, we use a combination of oscillating gradient and b-tensor encoding waveforms to extract more micro-structural information from the MRI signal. Advanced diffusion weightings can help to tease out the role of more specific microstructural features in the diffusion MRI signal.

Another key hardware improvement is the advent of field probes to measure time-dependent B0 field perturbations. This unlocks full flexibility in trajectory design. Traditionally, we are restricted to rectilinear trajectories so that the artifacts from field perturbations are more benign. We have been capitalizing on the flexibility offered by field probes by designing center-out trajectories that shorten echo-times and lead to substantial increases in signal strength.

Check out the pre-print Laterally Oscillating Trajectory for Undersampling Slices: LOTUS

Diffusion MRI has well‑known challenges such as geometric distortion, eddy currents, motion sensitivity, and long scan times. How will the next generation of applications alleviate or even amplify these issues?

These challenges are not going away anytime soon but we, as a community, have made great strides in addressing them. Moderate geometric distortions can be corrected quite effectively with approaches that estimate a B0 map, and strategies for mitigating the effects of eddy currents (including field probe-based methods) and motion have been developed. These methods have become quite robust for the single-shot approaches typically used for diffusion MRI.

In my view, the largest challenge moving forward is scan time. There is an ever-increasing number of more advanced diffusion MRI techniques proposed that offer improvements to microstructural specificity. However, these improvements are generally bought with scan time, making it very difficult to have a “one-size-fits-all” acquisition for diffusion MRI. We used to just be able to “add a diffusion sequence to the protocol”, but now one needs to choose the correct kind of diffusion sequence for the application. For example, specific questions about axons in white matter may be best served with ultrahigh b-value diffusion MRI, while oscillating gradients may be best for applications involving ischemia.

Highly efficient trajectories and 3D acquisitions are being explored to balance the desire for more specific data and short scan times, and great progress is being made. Simultaneous multi-slice techniques have been transformational by reducing scan times by factors of at least 2 to 3 on standard performance clinical scanners with minimal drawbacks; however, more improvements are necessary to translate new diffusion methods (which can still require 10’s of minutes to multiple hours of data collection) for practical applications.

What role do you see AI and data‑driven reconstruction playing in enabling these new applications?

There are many roles that AI can potentially play. In particular, it holds great potential to accelerate many processes. This includes image reconstructions, which traditionally can require more time than the acquisition itself when advanced trajectories are used. However, AI must be deployed in a way that takes full advantage of the physics we already understand, and great care must be taken to avoid hallucinations. To this end, tremendous progress has been made in physics-informed neural networks for image reconstruction.

What role do you see traditional reconstructions having as AI reconstructions advance.

It is likely that in the future all reconstructions will have at least some usage of AI. For now, however, traditional reconstructions have an essential role in validating the accuracy of AI recons. For example, a very safe approach to reap the benefits of AI would be for physics-informed AI to be used to give real-time reconstructions to the operator so that things like patient motion during the scan and the need for rescans can be assessed, and then once a more traditional reconstruction completes some time later it could be compared automatically to confirm no hallucinations occurred and add more training data.

There are many different techniques for fitting diffusion MRI data to obtain quantitative parameters. How does one go about choosing the fitting technique to use?

First of all, what you can fit is strongly determined by the type of data you acquire (b-values, diffusion weighting shapes, etc). Beyond that, different fitting approaches can be used for the exact same data. These can be split into two main classes: representations and models. Both classes have their own caveats and roles:

• Representations fit parameters empirically using a model that matches the data trends, which can be diagnostically useful. When there is a difference between patients and controls in a parameter obtained from a representation, the underlying biological mechanisms for that difference may be difficult to interpret. However, for diagnosis one just needs to know if a certain parameter change is associated with a condition, and the specific microstructural mechanisms are not as relevant. Moreover, certain representations can be acquired extremely efficiently, making them attractive from a scan-time perspective. 
• Microstructural models impose assumptions about the microstructural characteristics of the tissue in an effort to extract physiologically meaningful parameters, which give mechanistic insight. They have the potential to be extremely useful for improving our understanding of aging and disease since we can directly relate the results to specific biophysical features and processes. However, there is potential for model assumptions to be incorrect, and models must be rigorously validated and should ideally only be applied in tissue types and conditions where the assumptions are known to hold.

So, in summary, pick a model when you are looking for mechanistic insight, and choose one that has been validated for your application. Pick a representation when you have doubts about model validity or if you are primarily concerned about diagnosis. Sometimes it is appropriate to do both!

The Baron Lab at the Western University in London, ON