T2-PILOT: Optimized Trajectories for $T_2$ Mapping Acceleration

Naama Gavrielov, Tamir Shor, Alex M. Bronstein, Moti Freiman

Keywords: Cardiac MRI, $T_2$ Mapping, Trajectory Optimization and Reconstruction, Physics-Informed Deep-Learning

TL;DR: T2-PILOT reduces per-beat acquisition time for cardiac MRI by integrating the $T_2$ decay model into the joint optimization of trajectories and reconstruction, ensuring superior quantitative accuracy.

Registration Requirement: Yes

Abstract: Cardiac MRI $T_2$ mapping is essential for diagnosing myocardial pathologies, but prolonged acquisitions often extend beyond the diastolic rest phase, leading to motion artifacts and reduced reliability. While deep learning accelerates imaging via k-space undersampling, existing learned trajectories optimize for image reconstruction and neglect the underlying $T_2$ relaxation physics. We propose T2-PILOT, which jointly optimizes non-Cartesian k-space trajectories and $T_2$ map estimation by enforcing the $T_2$ decay model during training, with additional subject-specific test-time fine-tuning. On the CMRxRecon dataset, T2-PILOT outperforms both fixed and unconstrained reconstruction-guided trajectories. Under high undersampling, it improves image quality and quantitative $T_2$ accuracy while reducing per-beat acquisition time by 54\% (32 vs. 70 spokes), yielding a PSNR gain of 0.14 dB (35.06 vs. 34.92) and a $T_2$ map accuracy improvement of 0.67 dB (31.96 vs. 31.29). These results demonstrate that incorporating physics-based constraints into trajectory learning enables more accurate, robust, and clinically reliable accelerated $T_2$ mapping.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 35