Task-driven optimization of MRI sampling and IVIM parameter estimation

Abstract

Magnetic Resonance Imaging (MRI) plays a central role in non-invasive, high-resolution characterization of tissue microstructure. Among its modalities, diffusion-weighted imaging (DWI) supports estimation of intravoxel incoherent motion (IVIM) parameters, which capture both diffusion and perfusion processes. However, conventional IVIM pipelines rely on fixed acquisition patterns and stage-wise training, often yielding suboptimal accuracy under undersampling.

In this work, we propose a task-driven, end-to-end framework that couples learnable k-space sampling (Multi-LOUPE), image reconstruction, and tissue-specific IVIM parameter estimation. The pipeline integrates a tissue-weighted task loss and a dynamic multi-phase training strategy that progressively shifts focus from reconstruction fidelity to parameter accuracy, enabling gradient flow from IVIM supervision back to the sampling masks.

Experiments on the large-scale VICTRE breast phantom dataset demonstrate that our framework achieves balanced tissue-wise accuracy across tumor, vessel, and fat regions, matching or surpassing pattern-specific models while eliminating the need for retraining per sampling strategy. Additional analyses further confirm that the learned sampling masks capture complementary frequency information, explaining their robustness under undersampling.

Overall, this study shows that optimizing the entire flow—from what to measure in k-space to how to estimate quantitative parameters—outperforms stage-wise designs, offering a more generalizable and efficient pathway for accelerated quantitative MRI.