Author: Prabhu C. Acharya, Hassan Bagher-Ebadian, Stephen L. Brown, James R. Ewing, Mohammad M. Ghassemi, Benjamin Movsas, Farzan Siddiqui, Kundan S Thind 👨🔬
Affiliation: Michigan State University, Oakland University, Henry Ford Health 🌍
Purpose: Accurate T1 quantification using T One by Multiple Read Out Pulse (TOMROP) sequences is essential for physiological assessments in dynamic-contrast-enhanced (DCE) MRI and T1 mapping studies. Traditional multi-parametric fitting methods, such as Simplex, are widely used but computationally intensive and dependent on noise and pulse-sequence timing, leading to potential inaccuracies. This study introduces a novel equation of state that simplifies signal representation and enhances noise resilience, combined with an Artificial Neural Network (ANN), for rapid and accurate T₁ estimation from Look-Locker (LL) signals.
Methods: We derived a novel equation of state for LL signals to reduce its complexity and to simulate datasets with Rician noise, spanning 18 levels (6.9–20 dB SNR) for 36 excitations. A fully connected feedforward ANN architecture (36:10:1) optimized for regression tasks was trained, validated, and tested on 52,324 signals, adhering to 10-fold cross-validation. The trained-ANN was applied on the LL datasets of 17 rat brains acquired before and after DCE-MRI experiment. The ANN’s predictions were compared to those from Simplex fitting using correlation coefficients.
Results: The ANN achieved high predictive accuracy (97%, CI: 96–98%) in simulations and demonstrated strong agreement with Simplex fitting (r = 0.87, CI: 0.85–0.89) in rat brain studies. The ANN consistently provided rapid (~50 times faster than Simplex) and stable T1 estimates, even under high noise conditions, where Simplex fitting often struggles. Additionally, ANN-derived T1 maps showed superior spatial detail and alignment with literature-reported values.
Conclusion: This study introduces a novel equation of state integrated with an adaptive model for direct estimation of T1 from LL signals, effectively combining analytical modeling with machine learning. This approach offers faster, more accurate, and noise-resilient T1 quantification from LL pulse-sequences, with significant potential for DCE-MRI and other advanced T1 mapping applications.