1356639 - Deep learning-based non-rigid motion-corrected reconstruction for multi-contrast data

Multi-contrast and multiparametric myocardial tissue characterization plays an important role in the assessment of many cardiac diseases¹. However, reconstruction of motion-compensated multi-contrast 3D data is computationally expensive and time consuming. In this work, we propose to investigate the extension of previously introduced end-to-end deep learning (DL) technique for nonrigid motion-corrected reconstruction of undersampled free-breathing whole-heart coronary MRA² (MoCo-MoDL) for multi-contrast data. To investigate this, as a pilot study, we apply MoCo-MoDL to the multi-contrast data used in a previously described 3D joint T₁/T₂ mapping sequenece³ which includes inversion recovery, T₂ prepared, as well as non-prepared GRE acquisitions for in-phase and out-of-phase echoes, representing a wide range of contrasts.

Methods:

The proposed pipeline is shown in Fig.1. To evaluate the MoCo-MoDL for multi contrast data we use joint T₁/T₂ imaging³ with a 4x undersampled VD-CASPR trajectory⁴. This approach acquires four interleaved, ECG-triggered spoiled GRE volumes with 2-point bipolar Dixon encoding. Prior to the first and fourth volume an inversion recovery and T2 preparation pulse is performed, respectively.

Acquisitions were performed at 1.5T (Siemens Healthineers) on 58 subjects (34 patients and 24 healthy volunteers). Parameters: resolution=2mm², FA=8˚, TR=6.67ms, TE1/TE2=2.38/4.76ms, scan time ~9min. The data was pre-processed for each contrast by computing an undersampled zero-filled reconstruction for each respiratory bin (N_b=4) and a non-rigid motion-corrected reference image with HD-PROST reconstruction⁵. Zero-filled reconstructions and reference images for all contrasts were given to the network for training.

MoCo-MoDL² consists of a diffeomorphic respiratory motion estimation network in combination with a motion-informed model-based DL reconstruction network. The network was trained with the same parameters for each contrast using an ADAM optimizer with an initial learning rate of 2e^-4, which was reduced by half every 2000 iterations. The network was trained for 150 epochs with a training/testing split of 52/6 subjects.

Results:

An example of the learned multi-contrast MoCo-MoDL reconstruction for all eight contrasts and the respective HD-PROST reconstruction is shown in Fig. 2, showing good agreement for all contrasts. We report a reconstruction time of three sec per contrast for the proposed multi-contrast MoCo-MoDL while it takes around two hours for the joint HD-PROST reconstruction.

Conclusion:

The proposed multi-contrast MoCo-MoDL shows similar performance to HD-PROST while reducing reconstruction time by a factor of 300. These results show promise for future potential clinical applications such as the integration into the recently proposed motion compensated 3D joint T₁/T₂ mapping³. Future work should evaluate the proposed approach for 3D joint T₁/T₂ mapping and extend it to a joint reconstruction for all eight contrasts.