Accelerated EPG-based myocardial T1 mapping with PENGUIN

Abstract References Acknowledgements Accelerated EPG-based myocardial T1 mapping with PENGUIN INTRODUCTION: Model-based Deep Learning (DL) allows accelerating MRI reconstruction and quantitative mapping. Here, we propose a DL architecture that performs myocardial T1 mapping directly from accelerated k-space, which models the signal with the Extended Phase Graph (EPG) formulation, to allow more accurate quantification. The DL architecture is a modification of PhasE Graph sigNal and Gradients QUantitative Inference machiNe (PENGUIN)1, with the following novelties: (a) quantitative mapping is performed directly from undersampled k-space; (b) signal intensity curves are simulated not only for a range of T1 values, but also for a range of regular heart rate (HR) values. METHODS: PENGUIN (Figure 1a) combines a Recurrent Inference Machine2 with a dictionary of EPG simulated signal evolution curves that provides the network with a pre-calculated signal model. PENGUIN performs 𝐽=2 inference steps to obtain 𝐽 estimates of the T1 maps, considering the L1-norm loss function evaluated in the myocardium. Networks were trained for acceleration factors acc={4,8}, ADAM optimizer, learning rate=1e-4, for 300 epochs, 𝐶=64 and 𝐶=256 channels for acceleration factors 4 and 8, respectively. Data were obtained from MICCAI’s 2023 CMR reconstruction challenge3. T1 mapping was conducted following a 4-(1)-3-(1)-2 MOLLI sequence, short axis (SA) view only, FOV=360×307 mm2, spatial resolution=1.4×1.4 mm2, slice thickness=5.0 mm, TR=2.67 ms, TE=1.13 ms, partial Fourier=7/8, and GRAPPA factor of 2, on a 3T MAGNETOM Vida Siemens scanner. K-space data were undersampled retrospectively, following a radial k-t sampling trajectory with golden angle increments. EPG-based simulations were implemented for HR=[28,102] bpm, T1=0:1:2000 ms, T2=50 ms. Ground-truth T1 maps were obtained by performing a pattern recognition approach over the reconstructed, fully sampled signal intensity images, through dot product matching. Zero-filled (ZF) and Compressed Sensing (CS) reconstructions were performed to compare with PENGUIN. RESULTS & DISCUSSION: PENGUIN’s T1 maps (Figure 1b-c) achieved mean relative errors of 10.1±11.0% and 15.9±15.4%, and mean MSSIM scores of 0.999±0.001 and 0.998±0.001, for acc=4 and 8, respectively. PENGUIN is less resource-consuming than CS reconstruction, since the computational burden is moved to the preprocessing stage. Catarina N Carvalho1,2 *, Andreia S Gaspar2, Rita G Nunes2, Teresa M Correia1,3 1Center of Marine Sciences (CCMAR), Faro, Portugal; 2Institute for Systems and Robotics - Lisboa and Department of Bioengineering, Instituto Superior Técnico, Universidade de Lisboa, Lisbon, Portugal; 3School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom *[email protected] Figure 1 – (a) PENGUIN architecture for inference step 𝑗. The estimate 𝑝𝑗 is given to the dictionary to obtain the signal-intensity and corresponding derivative of each T1 value in 𝑝𝑗. The gradient of the negative log-likelihood Δ𝐿𝑗 is calculated, concatenated to 𝑝𝑗 and given as input to the network, which outputs the incremental update to the estimated maps. ℎ𝑗: memory vectors. (b) T1 maps of 5 distinct test subjects (rows), acc=4; ground truth (GT), Exponential Fitting, zero-filled (ZF), compressed sensing (CS), and PENGUIN. (c) T1 relative errors and MSSIM scores obtained for all testing images; * (p-value<0.05) and *** (p-value<0.001). 1. Carvalho C, Gaspar A, Nunes R, Correia T. Diving into Extended Phase Graph-based Deep Learning for accurate T2 mapping with PENGUIN. In proceedings of 2023 ISMRM & ISMRT Annual Meeting & Exhibition; 2. Putzky P, Welling M. Recurrent Inference Machines for Solving Inverse Problems. June 2017. 3. Wang C, Lyu J, Wang S, et al. CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI. Sci Data. 2024;11(1):687. a) b) c) This work was supported by: NVIDIA GPU hardware grant; “la Caixa” Foundation and FCT, I.P [LCF/PR/HR22/00533]; FCT (SFRH/BD/120006/2016, PTDC/EMD/EMD/29686/2017; UIDP/50009/2020), Programa Operacional Regional de Lisboa 2020 (LISBOA01-0145-FEDER-029686), LARSyS funding (DOI: 10.54499/LA/P/0083/2020,10.54499/UIDP/50009/2020,10.54499/UIDB/50009/2020. This research was supported by FCT through projects UIDB/04326/2020 (DOI:10.54499/UIDB/04326/2020), UIDP/04326/2020 (DOI:10.54499/UIDP/04326/2020) and LA/P/0101/2020 (DOI:10.54499/LA/P/0101/2020).