Presented at the first International Society for Tractography (IST) conference, 2025.

Work by Sanna Persson, Fabian Sinzinger, and Rodrigo Moreno.

Abstract

Diffusion-weighted imaging (DWI) acquisition is often constrained by scan time, producing sparsely and non-uniformly sampled Q-space. These limitations hinder accurate modelling of the diffusion signal and affect downstream tractography.

We propose representing the raw DWI signal with implicit neural representations (INRs): continuous, coordinate-based neural networks that provide a resolution-independent description of the orientation-distribution field rather than a fixed grid-sampled volume. Since INR-based DWI reconstructions only approximate the measured signal, we assess their effect on a practical downstream task — streamline tractography — to quantify when the continuous representation gains or loses relative to conventional reconstructions.