New MoE-dqINR framework cuts MRI reconstruction to 30 seconds

A research team led by Yinzhe Wu has introduced MoE-dqINR, a unified mixture-of-experts (MoE) framework for scan-specific dynamic and quantitative MRI reconstruction. The key innovation is separating shared spatial representation from state-dependent synthesis: a bank of spatial experts encodes reusable coordinate-dependent image content, while a routing network conditioned on acquisition states dynamically selects experts to form each frame or contrast state. This image-first architecture replaces monolithic spatiotemporal coordinate fields or explicit motion models, dramatically improving flexibility and computational efficiency. The framework integrates seamlessly with the multicoil MRI forward model and supports both dynamic (e.g., cardiac cine) and quantitative (e.g., T1/T2 mapping) acquisitions. In experiments, MoE-dqINR reduced per-scan optimization time to approximately 30 seconds—a 10x to 100x speedup over existing INR-based reconstructions that often require hundreds to thousands of seconds.

The practical impact is significant for clinical workflows: rapid, scan-specific reconstruction could enable real-time or near-real-time MRI without sacrificing image quality. By reusing shared spatial experts across acquisition states, the framework also generalizes better to unseen dynamics and provides a principled way to balance fidelity and efficiency. The authors demonstrate reconstructions from highly undersampled multicoil k-space data, showing robust recovery of spatiotemporal and quantitative maps. This work positions mixture-of-experts INRs as a new prior for medical imaging, unifying speed, flexibility, and state-of-the-art reconstruction quality. Future directions may include extension to 3D volumes and integration with generic sequence-based imaging models.