Related Experiment Video
Updated: Jun 11, 2026
10:17
Real-Time Cardiac Mapping with a Noninvasive Imageless Electrocardiographic Imaging System
Published on: April 11, 2025
Transformer networks enable fast and robust dictionary generation for multiparametric cardiac mapping with variable
Pauline Calarnou1, Amaury George1, Costa Georgantas1
1Department of Radiology, Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland.
Summary
A novel transformer network significantly accelerates cardiac multiparametric mapping by over 100-fold compared to traditional simulations. This deep learning approach maintains high accuracy and robustness, enabling potential routine in-line quantitative T1-T2 mapping.
Area of Science:
- Magnetic Resonance Imaging (MRI)
- Medical Imaging
- Artificial Intelligence in Medicine
Background:
- Cardiac multiparametric mapping typically uses dictionary matching, which is time-consuming due to the need for simulations with varying acquisition timings.
- This time constraint limits the clinical utility of in-line, real-time dictionary generation for cardiac mapping.
- Deep learning offers a potential solution to accelerate the dictionary generation process.
Purpose of the Study:
- To implement and evaluate a transformer network for accelerating cardiac multiparametric map reconstruction.
- To compare the transformer network's performance against a fully connected multi-layer perceptron (FC-MLP) and traditional Extended Phase Graph (EPG) simulations.
- To assess the clinical feasibility and potential for in-line deployment of the transformer-based mapping pipeline.
Main Methods:
- Developed and trained transformer and FC-MLP networks on synthetic timing intervals from EPG simulations.
- Quantified network performance using mean average percent error (MAPE) against EPG simulations, phantom correlation, and Bland-Altman analysis in 138 patients at 3T.
- Integrated the transformer network pipeline directly onto an MRI scanner for on-scanner inference.
Main Results:
- The transformer network achieved over 100-fold acceleration (inference in milliseconds vs. minutes for EPG simulations).
- Transformer demonstrated superior accuracy and robustness, with significantly lower MAPE for T1 and T2 mapping compared to FC-MLP, and minimal bias in Bland-Altman analysis.
- In vivo quantitative T1-T2 maps generated by the transformer were visually indistinguishable from EPG-derived maps, with very small biases.
Conclusions:
- The transformer network significantly accelerates cardiac multiparametric map reconstruction while maintaining accuracy comparable to EPG simulations and superior to FC-MLP.
- The network's robustness to variable acquisition timing and minimal bias support its clinical feasibility.
- Successful on-scanner integration highlights the potential for routine in-line deployment of quantitative T1-T2 mapping.