UNIFORM: A Unified Deep Learning Framework for Multi-organ and Multi-contrast MRI Reconstruction

George Yiasemis, Jonatan Ferm, Nikita Moriakov, Ritse Mann, Jan-Jakob Sonke, Jonas Teuwen

Keywords: Accelerated MRI Reconstruction, Deep Learning, Multimodal MRI, Zero-shot Self-supervised Learning

Abstract: MRI is an essential medical imaging modality, yet long acquisition times and organ-specific reconstruction methods often hinder clinical efficiency. In this paper, we propose training a unified deep learning framework (UNIFORM) for reconstructing undersampled multi-coil MRI data across diverse anatomical sites and multiple contrasts. Leveraging a state-of-the-art MRI reconstruction algorithm (vSHARP), UNIFORM was trained on diverse multi-coil $k$-space datasets, including knee, brain, prostate, and cardiac MRI. Evaluated across multiple acceleration factors ($2\times, 4\times, 6\times, 8\times$), it demonstrated robust performance in terms of quantitative evaluation. Additionally, UNIFORM supports zero-shot self-supervised learning (SSL), enabling effective reconstruction of unseen organs. Zero-shot SSL experiments were conducted on prospectively undersampled breast MRI acquisitions at high acceleration factors ($10\times, 17\times$), demonstrating improved anatomical detail and reduced noise compared to conventional zero-filling approaches. UNIFORM offers a promising avenue for clinically robust, accelerated multi-organ and multimodal MRI workflows.

Submission Number: 6