Accelerated Non-Contrast-Enhanced Three-Dimensional Cardiovascular Magnetic Resonance Deep Learning Reconstruction.

Background: Cardiovascular magnetic resonance (CMR) is a time-consuming, yet critical imaging method. In contrast, while rapid techniques accelerate image acquisition, these methods can also compromise image quality. Meanwhile, the effectiveness of Adaptive CS-Net, a vendor-supported deep-learning magnetic resonance (MR) reconstruction algorithm, for non-contrast three-dimensional (3D) whole-heart imaging using relaxation-enhanced angiography without contrast and triggering (REACT) remains uncertain.

Methods: Thirty participants were prospectively recruited for this study. Each underwent non-contrast imaging that included a modified REACT sequence and a standard 3D balanced steady-state free precession (bSSFP) sequence. The REACT data were acquired through six-fold undersampling and reconstructed offline using both conventional compressed sensing (CS) and an Adaptive CS-Net algorithm. Subjective and objective image quality assessments, as well as cross-sectional area measurements of selected vessels, were conducted to compare the REACT images reconstructed using Adaptive CS-Net against those reconstructed using conventional CS, as well as the standard bSSFP sequence. For a statistical comparison of image quality across these three image sets, the nonparametric Friedman test was performed, followed by Dunn's post-hoc test.

Results: The Adaptive CS-Net and CS-reconstructed REACT images exhibited superior image quality for pulmonary veins, neck, and upper thoracic vessels compared to the standard 3D bSSFP sequence. Adaptive CS-Net and CS reconstructed REACT images displayed significantly higher contrast-to-noise ratio (CNR) compared to those reconstructed using the 3D bSSFP sequence (all p-values < 0.05) for the left upper (5.40, 5.53, 0.97), left lower (6.33, 5.84, 2.27), right upper (5.49, 6.74, 1.18), and right lower pulmonary veins (6.71, 6.41, 1.26). Additionally, REACT methods showed a statistically significant improvement in CNR for both the ascending aorta and superior vena cava compared to the 3D bSSFP sequence.

Conclusions: The Adaptive CS-Net reconstruction for the REACT images consistently delivered superior or comparable image quality compared to the CS technique. Notably, the Adaptive CS-Net reconstruction provides significantly enhanced image quality for pulmonary veins, neck, and upper thoracic vessels compared to 3D bSSFP.