Synopsis

Keywords: Machine Learning/Artificial Intelligence, Motion Correction

Myocardial T₁ and T₂ mapping play an important role in the assessment of cardiovascular disease. 3D whole-heart joint T₁/T₂ water/fat mapping approaches have been recently proposed, however they require long reconstruction times. Recently a Machine learning based reconstruction was proposed for joint motion correction and motion corrected image reconstruction of undersampled free-breathing single contrast 3D coronary MR angiography. Here, we extend this approach for non-rigid motion-corrected reconstructions for multi-contrast data for joint T₁/T₂ mapping. The proposed approach achieves good agreement with reference techniques and comparable image quality to state-of-the-art methods albeit in 1200 times shorter reconstruction times.

Introduction

Myocardial T₁ and T₂ mapping play an important role in the assessment of cardiovascular disease (CVD)¹. Recently, a free-breathing high resolution, motion compensated 3D joint T₁/T₂ water/fat sequence has been proposed and validated in phantom and healthy subjects². To obtain the 3D whole-heart T₁ and T₂ maps, a reconstruction of eight intermediate motion-compensated volumes are needed, which is very time consuming. In this work, we propose to extend the previously introduced end-to-end deep learning (DL) technique initially applied to nonrigid motion-corrected reconstruction of undersampled free-breathing whole-heart coronary MRA³ for multi-parametric isotropic joint T₁/T₂ data. This allows a concurrent estimation of the motion fields and motion corrected reconstructions from the undersampled data, while increasing the speed and maintaining the quality of the reconstructions

Methods

The proposed pipeline is shown in Figures 1 and 2 for acquisition and reconstruction, respectively. The joint T₁/T₂ water/fat mapping research sequence² uses a 4x undersampled variable density cartesian trajectory³ and acquires four interleaved, ECG-triggered spoiled gradient-echo volumes with 2-point bipolar Dixon encoding. Prior to the first and fourth volume an inversion recovery and T₂ preparation pulse is used, whereas no preparation is used prior the second and third interleaves (Fig. 1).

The proposed DL-based reconstruction for multi-contrast undersampled free-breathing whole-heart data is extended from the end-to-end DL technique for nonrigid motion-corrected reconstruction of single contrast data (MoCo-MoDL)⁴. MoCo-MoDL consists of a diffeomorphic respiratory motion estimation network in combination with a motion-informed model-based deep learning reconstruction network. We adapted the network in order to process all eight contrasts, i.e. four interleaves and respective dual-echo signals (Fig. 1). We trained a network for each of the eight contrasts using an ADAM optimizer with an initial learning rate of 2x10^-4, which was reduced by half every 2000 iterations. The network was trained for 100 epochs with a training/testing split of 52/6. The same network architecture and hyper-parameters was used for each contrast.

After reconstruction, a water-fat separation algorithm is applied and a patient specific dictionary of pre-calculated T₁/T₂ combinations is matched against the measured water image signal evolution, resulting in the final T₁/T₂ maps². The precomputed T₁/T₂ combinations in the dictionary were simulated using the EPG formalism⁶ and are in the ranges T₁: 60:10:2000 ms, T₂: 4:2:100, 100:5:200, 200:10:300 ms.

Acquisitions were performed at 1.5T (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany) on 58 subjects (34 patients with suspected CVD and 24 healthy subjects). Main acquisition parameters were FA=8˚, TR = 6.67ms, TE1/TE2=2.38/4.76ms, 2 mm3 isotropic resolution and a subject specific mid-diastolic trigger-delay and acquisition window of ~100ms, resulting in a total scan time of nine minutes. The data was pre-processed for each contrast by computing a high quality refence image with a patch-based higher order low-rank (HD-PROST) reconstruction⁵. These high-quality respiratory bin images, together with undersampled respiratory bin images for all contrasts reconstructed using zero-filling, are inputted into the network.

Results

Reconstruction results for the proposed multicontrast MoCo-MoDL (MC MoCo-MoDL) are shown in Fig. 3 for all eight contrasts in comparison to state-of-the-art non-rigid motion-corrected HD-PROST(MoCo HD-PROST) reconstruction. Comparable visual image quality is achieved between the proposed approach and MoCo HD-PROST. 3D T₁/T₂ maps for the proposed MC MoCo-MoDL reconstruction in comparison to MoCo HD-PROST are shown in Fig. 4 for a representative subject. Mid-ventricular septum T₁ and T₂ values of 1171ms and 44ms, and 1021ms and 43ms were measured for MC MoCo-MoDL and MoCo HD-PROST, which were in good agreement with reference clinical 2D MOLLI and T₂ prepared bSSFP maps (840ms and 53ms respectively) (Fig. 5). Comparable visual image quality is achieved between the proposed approach and MoCo HD-PROST for T₂ mapping however remaining artefacts are observed with MC MoCo-MoDL for T₁ mapping. Reconstruction time per contrast was of 3s for the proposed MC MoCo-MoDL and of 2 hours for MoCo HD-PROST.

Discussion

This work is a proof of concept study and comprehensive tuning of the hyper parameters has not been performed yet. We believe that tuning the hyper parameters in the proposed network overall and per contrast can further improve the reconstruction quality, especially for T₁ mapping. Furthermore, a joint reconstruction of all contrasts with shared network parameters could improve the reconstruction quality further, and this will be investigated in future work.

Conclusion

We have proposed an end-to-end deep learning approach for non-rigid motion-corrected reconstructions for multi-dimensional 3D whole-heart joint T₁/T₂ mapping. The proposed approach achieves similar reconstruction quality as the non-rigid HD-PROST reconstruction used as reference, albeit with significantly shorter reconstruction times (1200 times faster). The reconstruction time is reduced to three seconds per contrast, which may make the approach promising for clinical application.

Acknowledgements

This work was supported by the following grants: (1) EPSRC P/V044087/1; (2) BHF programme grant RG/20/1/34802, (3) Wellcome/EPSRC Centre for Medical Engineering (WT 203148/Z/16/Z), (4) Millennium Institute for Intelligent Healthcare Engineering ICN2021_004, (5) FONDECYT 1210637 and 1210638, (6) IMPACT, Center of Interventional Medicine for Precision and Advanced Cellular Therapy, ANID FB210024.