Cardiac Diffusion Tensor Imaging (cDTI) is a cutting edge MRI technique that offers a glimpse into the detailed microstructure of the heart's muscle tissue. By tracking how water molecules move within the heart muscle, cDTI can map out the intricate structure of the tissue. This technique provides valuable information, which sheds light on how heart cells are orientated and how they work together, leading to new ways to diagnose and treat heart diseases [1].
Limitation of cDTI: The Motivation to Use Deep Learning
Despite the exciting potential of cardiac Diffusion Tensor Imaging in revealing the intricate microstructure of the heart muscle, several limitations hinder its widespread adoption in clinical practice.
One of the primary challenges lies in the need for scanning from multiple directions to accurately map the heart tissue. The process becomes even time-consuming with the need for multiple repetitions to enhance image quality.This requirement significantly prolongs the duration of MRI scans, imposing discomfort on patients, especially since it often necessitates holding one's breath during imaging.
Another significant obstacle is the inherent low signal-to-noise ratio (SNR) associated with single-shot acquisition techniques like single-shot echo planar imaging or spiral diffusion-weighted imaging. These techniques, while useful for fast scanning to mitigate motion artefacts from heartbeat and breathing, tend to produce lower quality images. Therefore, the trade-off for speed is several repetitions to achieve an acceptable level of accuracy in the DT estimation.
To speed up cDTI, researchers have been exploring two main strategies: reducing the number of images needed and using advanced techniques like deep learning to reconstruct high-quality images from fewer k-space measurement. Our focus here is on the latter, specifically how deep learning can help make cDTI clearer and faster.
Our Exploration: How Deep Learning Works for cDTI Reconstruction
Deep learning has emerged as a powerful technique for medical image analysis, capitalising on the nonlinear and complex nature of neural networks, and has been widely employed for MRI reconstruction, making it possible to produce clear and detailed images with subsampled measurements.
In this work, we investigated the application of deep learning-based methods for cDTI reconstruction including three representative deep learning-based models from algorithm unrolling models, enhancement-based models to emerging generative models in the cDTI dataset. The performance of these models was evaluated by the reconstruction quality assessment and the DT parameter assessment. Our pipeline of deep learning-based cDTI reconstructions is presented in Figure.1, which can be divided into four main steps: 1) data acquisition, 2) data pre-preprocessing, 3) deep learning-based cDTI reconstruction, and 4) data post-processing.
For data acquisition, retrospectively acquired cDTI data were acquired using a Siemens Skyra 3T MRI scanner and a Siemens Vida 3T MRI scanner (Siemens AG, Erlangen, Germany). A diffusion-weighted stimulated echo acquisition mode (STEAM) SS-EPI sequence with reduced phase field-of-view and fat saturation was used. We used 481 cDTI cases including 2 cardiac phases, that is, diastole (n = 232) and systole (n = 249), for the experiment section.
In the data pre-processing stage, all DWIs (b0, b150 and b600) were processed following the same protocol, including normalisation, zero-padding and simulation of k-space undersampling, as well as centre cropping.
For deep learning-based cDTI reconstruction, we implemented three representative deep learning-based models, including a Convolutional Neural Network (CNN)-based algorithm unrolling method, i.e., D5C5 [2], a CNN-based and conditional Generative Adversarial Network (GAN)-based method, i.e., DAGAN [3], and a Transformer-based and enhancement-based method, i.e., SwinMR [4].
In the data post-processing stage, we performed cDTI post-processing, following the protocol including: a) manual removal of low-quality DWIs; b) DWI registration; c) semi-manual segmentation for left ventricle (LV) myocardium; d) DT calculation using LLS fit; e) DT parameter calculation.
Figure.1 The data flow of our implementation for cardiac diffusion tensor imaging data. The whole procedure consists (A) data acquisition, (B) data pre-processing, (C) deep learning-based reconstruction, and (D) data post-processing.
Our Major Findings: Comparison of Deep Learning Models for cDTI Reconstruction
The results of our study were illuminating. We observed that each model showed distinct strengths in various acceleration factors (AFs) and evaluation metrics.
Regarding the fidelity of the reconstruction, D5C5 has shown superiority under the condition of a relative lower AF, while this superiority has been observed to disappear under the condition of a relative higher AF. This phenomenon is due to the use of data consistency modules (DC) in D5C5, which combine the k-space measurements with the CNN estimation to maintain consistency. With a relatively lower AF, a large proportion of information in the final output of D5C5 is provided by the DC module, whereas this proportion is significantly decreased in a relative higher AF reconstruction task (AF ×8). Therefore, this kind of unrolling-based methods with a DC module is more suitable for reconstruction at a relative lower AF.
For the perceptual score of the reconstructions, experiments have shown that SwinMR outperforms D5C5 and DAGAN on the metrics LPIPS and FID. However, even though the perceptual score has a high correlation with human observation, it is not always equivalent to a better reconstruction quality.
In the context of recovering the high-uncertainty region (green arrow in Figure.2), D5C5 successfully reconstructed the myocardium despite severe aliasing artefacts. DAGAN generated a reconstruction with low SNR and attempted to "in-paint" the missing myocardium based on its prior knowledge influenced by the adversarial learning process. SwinMR preserved most of the myocardium information, but the reconstruction was heavily affected by "fake" details (hallucination).
Hallucination is usually defined as artefacts or incorrect features that occur due to the prior that cannot be produced from the measurements [5]. Through empirical observations, we noted that such false details became more pronounced in challenging tasks with high ambiguity levels, or when utilising a robust "strong-prior" model such as Transformers (SwinMR) or generative models (DAGAN).
What's Next: Limitations and Future Directions
Our studies have revealed that there are still critical limitations when directly applying these general MRI reconstruction methods to cDTI reconstruction.
First, a notable limitation is the absence of specific restrictions on diffusion within the models employed. Traditional loss functions used in DAGAN, D5C5, and SwinMR focus predominantly on image and perceptual quality without accounting for specific diffusion information. Incorporating diffusion-specific constraints, such as physics-informed loss functions or embedding physical properties like b-values and diffusion directions into the training process, could significantly enhance model performance by aligning it more closely with the underlying physiological processes.
Second, our findings underscore a trade-off between perceptual performance and quantitative performance. The arrangement of myocardial fibres revealed by cDTI is influenced by the pixel intensity of DWIs, which reflects the actual physiological conditions. However, the models examined in this research were initially intended for structural MRI. In particular, DAGAN and SwinMR focused more on "perceptual-similarity", which can be seen as the distance in latent space. The perceptual loss used for training explicitly constrains the distance in latent space, while the adversarial learning mechanism implicitly minimises a statistical gap between two distributions in latent space. Therefore, future studies should focus on enhancing pixel-wise fidelity rather than perceptual similarity, or on preventing the introduction of artificial information.
Lastly, we encountered a gap between the conventional methods of evaluating diffusion tensor quality and the actual quality of cDTI reconstructions. Current evaluation metrics, such as the global mean value of diffusion parameters, are not always accurate or sensitive enough to evaluate the diffusion tensor quality. This limitation is illustrated in Figure.3, where traditional metrics suggested the equivalence between reconstruction results and reference data, despite clear visual evidence to the contrary. Moving forward, a more nuanced approach to assessment is necessary: one that goes beyond global averages to consider the spatial distribution of diffusion properties or leverages downstream task assessments, such as pathology classification, to provide a more holistic view of reconstruction quality.
In conclusion, our research has explored the utilisation of deep learning techniques to speed up cDTI reconstruction, showing great potential for enhancing the incorporation of cDTI into standard clinical procedures. The outcomes indicate that these models can be efficiently applied in clinical settings at acceleration factors of ×2 and ×4, with SwinMR being the preferred choice. Nevertheless, when operating at an acceleration factor of ×8, the performance of all models is still limited, necessitating further research to enhance their effectiveness at higher acceleration factors.
References
1. Ferreira, P. F. et al. In vivo cardiovascular magnetic resonance diffusion tensor imaging shows evidence of abnormal myocardial laminar orientations and mobility in hypertrophic cardiomyopathy. Journal of Cardiovascular Magnetic Resonance 16, (2014).
2. Schlemper, J., Caballero, J., Hajnal, J. V., Price, A. N. & Rueckert, D. A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction. IEEE Transactions on Medical Imaging 37, 491–503 (2018).
3. Yang, G. et al. DAGAN: Deep De-Aliasing Generative Adversarial Networks for Fast Compressed Sensing MRI Reconstruction. IEEE Transactions on Medical Imaging 37, 1310–1321 (2018).
4. Huang, J. et al. Swin transformer for fast MRI. Neurocomputing 493, 281–304 (2022).
5. Bhadra, S., Kelkar, V. A., Brooks, F. J. & Anastasio, M. A. On Hallucinations in Tomographic Image Reconstruction. IEEE Transactions on Medical Imaging 40, 3249–3260 (2021).
Please sign in or register for FREE
If you are a registered user on Research Communities by Springer Nature, please sign in