Related Experiment Video
Updated: Aug 5, 2026
07:17
Leveraging Virtual Reality for Immersive Segmentation and Analysis of Cryo-Electron Tomography Data
Published on: January 24, 2025
High Resolution Isotropic 'Pseudo' 3D Cine imaging with Automated Segmentation using Concatenated 2D Real-time
Mark Wrobel1, Tina Yao1, Ruaraidh Campbell1
1UCL Centre for Translational Cardiovascular Imaging, University College London, 20c Guilford St, London, UK, WC1N 1DZ.
Summary
Deep learning creates pseudo 3D-cine cardiovascular magnetic resonance (CMR) images from free-breathing real-time scans. This novel technique significantly speeds up clinical CMR by reducing scan times while maintaining diagnostic quality.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Cardiovascular Imaging
Background:
- Conventional cardiovascular magnetic resonance (CMR) uses 2D breath-hold (BH) cine imaging for function and static 3D whole-heart imaging for anatomy.
- Pediatric and congenital heart disease imaging presents unique challenges for standard CMR techniques.
Purpose of the Study:
- To develop and validate a Deep Learning (DL) method for creating isotropic, fully segmented 3D-cine datasets from 2D free-breathing real-time CMR images.
- To assess the feasibility of significantly reducing CMR scan times without compromising diagnostic accuracy.
Main Methods:
- Four DL models were trained for interslice signal and respiratory correction, super-resolution, and segmentation of cardiac structures and great vessels.
- Prospective sagittal stacks of real-time cine images from 20 patients were converted into segmented, pseudo 3D-cine data.
- Quantitative metrics (ventricular volumes, vessel diameters) and image quality were compared to reference-standard BH cine and 3D whole-heart imaging.
Main Results:
- The DL method successfully transformed real-time cine data into pseudo 3D-cines with offline reconstruction in under 1 minute.
- No significant biases were observed in left and right ventricular volume metrics, showing reasonable agreement with reference standards.
- Adequate diagnostic image quality was achieved for the DL pseudo-3D-cine data, outperforming unprocessed 2D real-time data.
Conclusions:
- Deep learning enables the creation of pseudo 3D-cine data from concatenated 2D real-time cine images, offering fully segmented data in under a minute.
- The method's short acquisition and reconstruction times, combined with open-source model availability, facilitate widespread clinical adoption.
- The strong agreement with reference-standard imaging suggests this technique can significantly accelerate CMR procedures in clinical practice.
Keywords:
3D-cineCongenital heart diseaseDeep-LearningFree-breathingNon-contrastPediatric cardiologyReal-time
