TL;DR
This paper introduces a U-Net based deep learning model that can reconstruct cardiac diffusion tensor imaging (cDTI) parameters from significantly reduced scan times, aiming for a single breath-hold. This approach shows promise in overcoming the current technical limitations of cDTI, making it more feasible for clinical use.
What Problem Does This Paper Address?
Cardiac diffusion tensor imaging (cDTI) is a powerful technique for characterizing the microstructure of the heart's muscle. However, acquiring high-quality cDTI data is challenging due to the inherently low signal-to-noise ratio (SNR) of the required diffusion-weighted images. To compensate for this, current protocols often require multiple repetitions, leading to long scan times and necessitating numerous breath-holds. This is particularly problematic for cardiac imaging, where patient motion and breath-holding can be difficult, especially for patients with cardiac conditions. The goal of this research is to drastically reduce the number of breath-holds required for cDTI, making the technique more practical and accessible for clinical applications.
Methods
The core of this work is a U-Net convolutional neural network (CNN) architecture. Unlike traditional methods that fit diffusion tensor models to acquired image data, this U-Net is trained to directly predict the diffusion tensor components from a reduced set of diffusion-weighted images. This is a "fitting-free" approach. The U-Net takes a small number of diffusion-weighted images (corresponding to 5, 3, or even 1 breath-hold) as input and outputs the six components of the diffusion tensor. These components are then used to calculate standard cDTI parameters like Fractional Anisotropy (FA), Mean Diffusivity (MD), Helix Angle (HA), and Sheetlet Angle (E2A). The training process uses a mean absolute error loss function, aiming to minimize the difference between the U-Net's predicted tensor components and those derived from a full, high-quality dataset.
Datasets
The study utilized a retrospective dataset comprising 744 cardiac diffusion tensor imaging (cDTI) scans from both healthy volunteers (n=197) and cardiac patients (n=547). These patients represented a variety of conditions including amyloidosis, dilated cardiomyopathy (DCM), Fabry's disease, hypertrophic cardiomyopathy (HCM), acute myocardial infarction (MI), Marfan's syndrome, and recovered DCM (rDCM). The data was acquired on Siemens Skyra and Vida 3T MRI scanners using a diffusion-weighted stimulated echo acquisition mode (STEAM) single-shot echo-planar imaging (EPI) sequence. For evaluation, the researchers also prospectively acquired data from five healthy volunteers, with scans taken at five equidistant slices from base to apex of the left ventricle.
Key Insights and Ideas
- Fitting-Free Prediction: The U-Net bypasses the complex and time-consuming tensor fitting process, directly learning the relationship between raw diffusion-weighted images and the final tensor parameters. This is a significant departure from conventional methods.
- Spatial Pattern Learning: Unlike voxel-wise fitting, the U-Net can leverage spatial patterns within the cardiac anatomy, potentially leading to more robust parameter estimation, especially in challenging cases.
- Significant Breath-Hold Reduction: The model demonstrates the feasibility of obtaining clinically useful cDTI information from as few as 1-3 breath-holds, a dramatic reduction from the typical 12+ breath-holds.
- U-Net Outperforms LLS at Reduced Data: For datasets acquired with 3 or 1 breath-hold, the U-Net model consistently produced lower pixel-wise errors compared to the conventional linear least-squares (LLS) tensor fitting method when using the same reduced data.
Key Results
- Reduced Scan Time: The U-Net approach enables scan time reductions of approximately 60% for 5 breath-holds and up to 90% for 1 breath-hold acquisitions.
- Improved Accuracy at Low Breath-Holds: For 3 and 1 breath-hold datasets, the U-Net significantly outperformed the LLS method in terms of pixel-wise errors across all measured parameters (FA, MD, HA, E2A).
- Global Parameter Bias: While pixel-wise errors were lower for the U-Net at reduced breath-holds, both LLS and U-Net methods with reduced datasets showed some bias and small but significant differences from the reference results for global mean/median values. Notably, LLS:MD 5BH (P=0.38) and MD 3BH (P=0.09) showed no significant difference from the reference.
- Smoother Parameter Maps: The U-Net predictions resulted in smoother parameter maps, which, while reducing some fine detail, also appeared to mitigate artifacts, particularly in lower breath-hold scenarios.
- Clinical Utility Demonstrated: In a cohort of acute myocardial infarction patients, the U-Net was able to identify infarcted regions based on diffusivity differences, although the contrast was reduced with fewer breath-holds.
Limitations and Future Work
- Smoothing Effect: The U-Net's tendency to produce smoother parameter maps might lead to a loss of fine detail and reduced contrast in abnormal areas, especially with very limited data. Future work could explore residual dense blocks or other techniques to improve high-frequency feature reconstruction.
- Generalizability: The model was trained and tested on data from specific scanner models and protocols. Its generalizability to other scanners and protocols is currently unknown and requires further investigation.
- Limited Training Data for Specific Cohorts: The small number of training samples for certain patient groups (e.g., Fabry, hypertensive DCM, Marfan) means the U-Net's performance for these specific conditions is uncertain. Expanding the training dataset is a key future direction.
- Uncertainty Estimation: The authors highlight the importance of quantifying model uncertainty, especially when dealing with limited data or potential artifacts. Future work could incorporate methods like deep ensembling to estimate predictive uncertainty.
- Prospective Validation: While some prospective data was acquired, the statistical analysis was limited due to the small sample size. Further prospective validation is needed.
- Advanced CNN Architectures: The authors suggest exploring more advanced CNN designs, such as conditional generative adversarial networks (cGANs) and incorporating residual dense blocks, to further enhance reconstruction quality.
Summary updated