VentrEX: An Anatomically Guided Deep Learning Pipeline for Ventricular Segmentation in Cine Cardiac MRI.
Authors
Affiliations (2)
Affiliations (2)
- School of Engineering, Computer Science and Artificial Intelligence, Long Island University, Brooklyn, NY 11201, USA.
- Unidad de Cirugía Cardiovascular, Hospital General de Niños "Dr. Pedro de Elizalde", Buenos Aires C1270AAN, Argentina.
Abstract
Automated segmentation of the left and right ventricles (LVs and RVs) in cine cardiac MRI (CMR) underpins reliable volumetry and mass estimation. However, papillary muscles and trabeculae (PM/T) introduce clinically meaningful variability and exacerbate cross-dataset domain shift. We present VentrEX, an anatomically guided pipeline. The core segmenter, VentrEX-Seg, is a 3D encoder-decoder with parallel channel-spatial attention and a Transformer bottleneck. Training is performed exclusively on ACDC. A lightweight PM/T module automatically extracts papillary and trabecular burdens and standardizes cavity volumes. External evaluation is zero-shot (no fine-tuning) on Sunnybrook (LV) and MM-WHS MRI (RV). We report Dice, HD95 (mm); for volumetry, we use Bland-Altman analyses (LV and RV volumes). Attention/Grad-CAM visualizations support interpretability. On ACDC, VentrEX achieved higher Dice and lower boundary error than U-Net, nnU-Net, CBAM, and VentrEX-Seg. Zero-shot performance was preserved externally (e.g., Sunnybrook LV Dice 0.9053, HD95 4.95 mm; MM-WHS RV Dice 0.9236, HD95 6.61 mm). Patient-level Bland-Altman analyses characterized LV and RV volumetric agreement. Qualitative overlays and 3D reconstructions showed fewer PM/T "leaks" and anatomically plausible borders across ED/ES. Single-source training with dual zero-shot external tests demonstrates robustness under domain shift. The combination of parallel attention and a Transformer bottleneck enables accurate, transparent cine-CMR segmentation across datasets.