A deep learning pipeline accurately reconstructed 3D cardiac anatomy from sparse CMR with 50-70% of slices removed, achieving 3D Dice similarity coefficients of 0.95 for LV and 0.93 for RV.
Does a deep learning pipeline accurately reconstruct 3D whole-heart anatomy from sparse 2D CMR images in patients with repaired Tetralogy of Fallot?
A deep learning pipeline can accurately reconstruct 3D whole-heart anatomy from sparse CMR data in repaired Tetralogy of Fallot, potentially reducing required scan times while maintaining diagnostic accuracy.
BACKGROUND: Accurate assessment of 3D four chamber cardiac anatomy is essential for managing repaired Tetralogy of Fallot (rToF), yet standard cardiac magnetic resonance (CMR) protocols acquire sparse 2D slices with anisotropic resolution, inter-slice inconsistencies, and motion artifacts from patient movement and incomplete acquisitions. As CT provides isotropic 3D whole-heart segmentations, it can be used to train deep learning models to reconstruct dense 3D anatomies from sparse CMR-like slices simulated from these volumes, bridging the gap between CMR's clinical accessibility and CT's spatial resolution. We aimed to develop and validate such a deep learning pipeline for comprehensive 3D whole-heart reconstruction from routine 2D cine short and long axis CMR images in rToF patients. METHODS: A two-stage deep learning pipeline was developed to reconstruct cardiac anatomy from sparse two-dimensional cine (2D+time) CMR images. A UNet segmentation network was trained on 152 and tested on 19 held-out rToF patients from multiple centres. A label completion network reconstructed dense 3D geometries from sparse segmentations using 1,715 CT segmentations augmented with synthetic right ventricle (RV) myocardium and simulated ToF-specific views. The approach was validated using simulated slices from 18 3D non-contrast enhanced magnetic resonance angiography scans and clinical CMR scans. To simulate incomplete acquisitions encountered in paediatric clinical practice - where scans are frequently abandoned prematurely or individual views are missing due to patient motion or non-compliance - reconstruction performance was evaluated by randomly removing 50-70% of short-axis slices prior to 3D geometry reconstruction. RESULTS: With 50-70% of short-axis slices randomly removed, reconstruction achieved 3D Dice similarity coefficient (DSC) of 0.95 (left ventricle (LV)), 0.93 (right ventricle (RV)), 0.89 (LV myocardium), 0.89 (left atrium), 0.79 (right atrium), and 0.74 (RV myocardium). Chamber volumes and masses were accurately recovered: LV volume -0.27±1.60mL, RV volume 3.91±4.98mL, LA volume -2.32±3.33mL, LV mass -1.87±2.08g, RV mass 2.47±5.6g. On test CMR cases (n=19) with motion artifacts, 2D-DSC were 0.95±0.02 (LV cavity), 0.91±0.02 (RV cavity), 0.95±0.06 (LA cavity), 0.91±0.09 (RA cavity), 0.70±0.06 (RV myocardium) and 0.88±0.03 (LV myocardium), meeting or exceeding inter-observer variability. The network accurately reconstructed geometries throughout the cardiac cycle despite training only on diastasis hearts. CONCLUSIONS: This pipeline enables anatomically accurate cardiac anatomy from sparse CMR data, potentially reducing scan times while maintaining diagnostic accuracy.
Mauger et al. (Mon,) conducted a other in Repaired Tetralogy of Fallot (rToF) (n=171). Deep learning pipeline for 3D whole-heart reconstruction vs. Ground truth segmentations was evaluated on 3D Dice similarity coefficient (DSC) and 2D-DSC for cardiac chambers. A deep learning pipeline accurately reconstructed 3D cardiac anatomy from sparse CMR with 50-70% of slices removed, achieving 3D Dice similarity coefficients of 0.95 for LV and 0.93 for RV.