Motion correction for native myocardial T₁ mapping using self‐supervised deep learning registration with contrast separation

Abstract

In myocardial T 1 mapping, undesirable motion poses significant challenges because uncorrected motion can affect T 1 estimation accuracy and cause incorrect diagnosis.In this study, we propose and evaluate a motion correction method for myocardial T 1 mapping using self-supervised deep learning based registration with contrast separation (SDRAP). A sparse coding based method was first proposed to separate the contrast component from T 1 -weighted (T1w) images. Then, a self-supervised deep neural network with cross-correlation (SDRAP-CC) or mutual information as the registration similarity measurement was developed to register contrast separated images, after which signal fitting was performed on the motion corrected T1w images to generate motion corrected T 1 maps. The registration network was trained and tested in 80 healthy volunteers with images acquired using the modified Look-Locker inversion recovery (MOLLI) sequence. The proposed SDRAP was compared with the free form deformation (FFD) registration method regarding (1) Dice similarity coefficient (DSC) and mean boundary error (MBE) of myocardium contours, (2) T 1 value and standard deviation (SD) of T 1 fitting, (3) subjective evaluation score for overall image quality and motion artifact level, and (4) computation time. Results showed that SDRAP-CC achieved the highest DSC of 85.0 ± 3.9% and the lowest MBE of 0.92 ± 0.25 mm among the methods compared. Additionally, SDRAP-CC performed the best by resulting in lower SD value (28.1 ± 17.6 ms) and higher subjective image quality scores (3.30 ± 0.79 for overall quality and 3.53 ± 0.68 for motion artifact) evaluated by a cardiologist. The proposed SDRAP took only 0.52 s to register one slice of MOLLI images, achieving about sevenfold acceleration over FFD (3.7 s/slice).

DeepFittingNet: A deep neural network‐based approach for simplifying cardiac T1 and T2 estimation with improved robustness

“…For comprehensive CMR myocardium tissue examination in clinical practice, generally including pre−/post‐contrast T 1 and T 2 mapping, although a dedicated deep‐learning fitting model could be built for each T 1 or T 2 mapping sequence, the proposed DeepFittingNet in the multi‐tasking manner could further simplify the post‐processing because there is no need to prepare models for different tasks (e.g., inversion‐recovery T 1 mapping and T 2 ‐prep bSSFP) or select models for specific applications (e.g., MOLLI5(3)3 for pre‐contrast T 1 and MOLLI4(1)3(1)2 for post‐contrast T 1 ). In the future, Pixel‐wise DeepFittingNet could be integrated into the workflow of reconstruction, quality control, motion correction, and segmentation, to further simplify and achieve fully automatic processing for myocardium tissue characterization 31‐33 …”

“…In the future, Pixel-wise DeepFittingNet could be integrated into the workflow of reconstruction, quality control, motion correction, and segmentation, to further simplify and achieve fully automatic processing for myocardium tissue characterization. [31][32][33] Previous studies have shown that physics/model-driven unrolling iterative deep-learning method can improve performance and interpretability. 34 In this study, we repeated the last two layers of FCNN of DeepFittingNet and utilized the difference between the measured and synthesized signals to simulate the procedure of minimizing residual in the conventional fitting.…”

Artificial Intelligence in Cardiovascular CT and MR Imaging

“…AI algorithms can also be used to reduce artefacts and to improve the accuracy of native myocardial value estimation. Li et al [ 87 ] evaluated a motion correction method for myocardial T1 mapping using self-supervised deep learning-based registration with contrast separation (SDRAP). Results showed that the AI algorithm achieved the highest DSC of 85.0 ± 3.9% and the lowest mean boundary error (MBE) of 0.92 ± 0.25 mm among the methods compared.…”

Motion-compensated T1 mapping in cardiovascular magnetic resonance imaging: a technical review

“…Motion correction and mapping took about 30 s per 2D slice. Li and Wu et al also showed a self-supervised registration method with contrast separation to improve motion characterization in mapping ( 51 ). The proposed deep learning registration approach was implemented on a GPU and shortened the computation time from 3.7 to 0.5 s. Additionally, an experienced cardiologist evaluated image quality of maps and images, finding an increase in image quality and reduction in motion artifacts when compared to standard motion correction methods.…”

“…Deep learning methods have also been explored for 2D T 1 mapping: Gonzales et al used T 1 w images to train a coarse-to- (51). The proposed deep learning registration approach was implemented on a GPU and shortened the computation time from 3.7 to 0.5 s. Additionally, an experienced cardiologist evaluated image quality of T 1 maps and T 1 w images, finding an increase in image quality and reduction in motion artifacts when compared to standard motion correction methods.…”