2025/07 - Joint image reconstruction and segmentation of real-time cardiovascular magnetic resonance imaging in free breathing using a model based on disentangled representation learning [Journal Club-Joint image reconstruction and segmentation of real-time cardiovascular magnetic resonance imaging in free breathing using a model based on disentangled representation learning

Joint image reconstruction and segmentation of real-time cardiovascular magnetic resonance imaging in free breathing using a model based on disentangled representation learning

Back to course

Video Transcription

Video Summary

The session introduced a JCMR Journal Club discussion on a paper about undersampled real-time cardiac cine MRI using disentangled representation learning for joint reconstruction and segmentation. The presenters explained that conventional cine MRI is reliable but slow, ECG-dependent, and vulnerable to breath-hold failure and arrhythmia. Their goal was to enable faster, free-breathing imaging with automatic segmentation.<br /><br />They developed a spiral balanced SSFP sequence and an X-SDNet model based on disentangled representation learning. The network jointly reconstructed images and produced segmentations of the left ventricle, right ventricle, and myocardium. Training combined simulated data and in-house acquisitions from healthy volunteers, plus validation in patients with atrial fibrillation. Compared with standard reconstruction methods, X-SDNet reduced undersampling artifacts and produced smooth, frame-by-frame segmentations without temporal jitter.<br /><br />Reader studies found image quality comparable to conventional Cartesian cine in healthy subjects and better temporal dynamics in arrhythmic patients. Ejection fraction measurements matched the reference well, with only a small bias. The presenters noted remaining limitations, especially small sample size and less robust right-ventricular segmentation. They also discussed future work toward higher temporal resolution and possible on-scanner implementation.

Keywords

cardiac cine MRI

undersampled reconstruction

disentangled representation learning

joint reconstruction and segmentation

X-SDNet

spiral balanced SSFP

free-breathing imaging

atrial fibrillation

ejection fraction measurement