In his project, Jesús tackled one of the key challenges of brain diffusion MR imaging at 7T: susceptibility-induced geometric distortions. Although 7T diffusion imaging allows for superior signal-to-noise ratio, it makes these distortions more severe, producing anatomically unreliable images. Existing tools work well enough at conventional field strengths, but at 7T there isn´t even a proper gold standard method. Additionally, head motion along with eddy-current-induced distortions misalign the series of images, biasing downstream analyses.

His approach consisted of a fully unsupervised, two-stage physics-informed deep learning framework that jointly corrects head motion, eddy-current, and susceptibility-induced distortions through image registration using a structural MR scan as reference. Unlike conventional methods, it doesn't require reverse phase-encoding acquisitions. Hard to evaluate rigorously in the absence of a ground truth, but the results were very promising.

But what if we went one step further back and solved the problem directly in the raw MRI signal domain, k-space, rather than in image space? That's exactly what he will be be doing next.

At DRCMR, we are very excited to continue our work with Jesús Díaz Pereira in his DDSA - Danish Data Science Academy funded PhD project, tackling challenges of high-resolution, distortion-free diffusion MRI. This project beautifully combines basic MR physics and advanced deep learning methods for the benefit of neuroscience applications, including in multiple sclerosis.