Related Experiment Video
Updated: Jul 1, 2025
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
22.5K
Deep learning-based diffusion tensor cardiac magnetic resonance reconstruction: a comparison study
Jiahao Huang1,2,3, Pedro F Ferreira4,5, Lichao Wang4,6
1National Heart and Lung Institute, Imperial College London, London, SW7 2AZ, UK. j.huang21@imperial.ac.uk.
Scientific Reports
|March 7, 2024
Summary
Deep learning models enhance in vivo cardiac diffusion tensor imaging (cDTI) reconstruction for clinical use. SwinMR is recommended for acceleration factors up to ×4, but higher factors require further development.
Area of Science:
- Cardiovascular Imaging
- Medical Image Reconstruction
- Artificial Intelligence in Medicine
Background:
- In vivo cardiac diffusion tensor imaging (cDTI) is a valuable MRI technique for assessing myocardial microstructure and cardiac function.
- Clinical adoption of cDTI is hindered by technical challenges like low signal-to-noise ratio and long scan times.
- Deep learning-based reconstruction offers a potential solution to accelerate cDTI acquisition.
Purpose of the Study:
- To investigate and implement three deep learning-based MRI reconstruction models for cDTI.
- To evaluate the performance of these models regarding reconstruction quality, diffusion tensor parameter accuracy, and computational cost.
- To determine the feasibility of using these models for clinical cDTI at various acceleration factors (AF).
Main Methods:
- Implementation of three distinct deep learning MRI reconstruction models.
- Assessment of reconstruction quality using objective metrics.
- Evaluation of diffusion tensor parameter accuracy and map quality at AF ×2, ×4, and ×8.
- Analysis of computational efficiency for each model.
Main Results:
- Deep learning models are suitable for clinical cDTI at AF ×2 and ×4.
- The D5C5 model demonstrated superior reconstruction fidelity, while SwinMR offered higher perceptual scores.
- At AF ×2 and ×4, diffusion tensor parameters showed no significant statistical difference from the reference, with acceptable map quality.
- Model performance degraded significantly at AF ×8, with limited parameter recovery and potential for misleading results.
Conclusions:
- Deep learning reconstruction models, particularly SwinMR, show promise for accelerating clinical cDTI at AF ×2 and ×4.
- Current models are not yet ready for clinical application at higher acceleration factors (AF ×8) due to performance limitations.
- Further research and development are needed to optimize deep learning models for higher acceleration factors in cDTI.
