Deep learning motion correction of quantitative stress perfusion cardiovascular magnetic resonance

BACKGROUND: Quantitative stress perfusion cardiovascular magnetic resonance (CMR) is a valuable tool for assessing myocardial ischemia. Motion correction is a crucial step in automated quantification pipelines, especially for high-resolution pixel-wise mapping. Established methods for motion correction, based on image registration, are computationally intensive and sensitive to changes in image acquisitions, necessitating more efficient and robust solutions. METHODS: This study developed and evaluated an unsupervised deep learning-based motion correction pipeline. Based on a previously described approach, it corrects motion in three steps while using (robust) principal component analysis to mitigate the effects of the dynamic contrast. The time-consuming iterative registration optimizations are replaced with an efficient one-shot estimation by trained deep learning models. The pipeline aligns the perfusion series and includes auxiliary images series as follows: the low-resolution, short-saturation preparation time arterial input function series and the proton density-weighted images. The deep learning models were trained and validated on multi-vendor data from 201 patients, with 38 held out for independent testing. The performance was evaluated in terms of the temporal alignment of the image series and the derived quantitative perfusion values in comparison to a previously established optimization-based registration approach. RESULTS: The deep learning approach significantly improved temporal smoothness of time-intensity curves compared to the previously published baseline (p<0.001). Temporal alignment of the myocardium (based on automated segmentations) was similar between methods and significantly improved for both as compared to before registration (mean (standard deviation) Dice = 0.92 (0.04) and Dice = 0.91 (0.05) (respectively) vs Dice = 0.80 (0.09), both p<0.001). Quantitative perfusion maps were also smoother, indicating a reduction of motion artifacts, with a median (interquartile range) standard deviation of 0.52 (0.39) ml/min/g in myocardial segments, than before motion correction and improved compared to the baseline method (0.55 [0.44] mL/min/g). Processing time was reduced by a factor of 15 for a representative image series using the deep learning approach in comparison to the iterative method. CONCLUSION: The deep learning approach offers faster and more robust motion correction for stress perfusion CMR, improving accuracy for the dynamic contrast-enhanced data and the auxiliary images. It was trained with multi-vendor data and different acquisition sequence implementations, so, as well as enhancing efficiency and performance, it could facilitate broader clinical use of quantitative perfusion CMR.