AID-DTI: Accelerating High-fidelity Diffusion Tensor Imaging with Detail-preserving Model-based Deep Learning

TL;DR

The paper addresses the challenge of obtaining high-fidelity DTI metric maps from highly sparse q-space data by proposing AID-DTI, a model-based deep learning framework that combines a novel SVD-based regularizer with a Nesterov-based adaptive hyperparameter optimization to learn from six diffusion directions. The method uses a flexible MESC-SD unfolded backbone and enforces consistency in the dominant singular subspaces of the predicted and ground-truth parameter maps, while dynamically tuning the regularization weight. On the Human Connectome Project dataset, AID-DTI outperforms conventional diffusion tensor fitting and several deep-learning baselines in MSE, SSIM, and PSNR, demonstrating finer detail preservation and robustness to noise. The work suggests AID-DTI can enable fast, high-fidelity DTI with minimal q-space data, with potential clinical applicability and avenues for extension to other diffusion models and multi-parametric imaging.

Abstract

Deep learning has shown great potential in accelerating diffusion tensor imaging (DTI). Nevertheless, existing methods tend to suffer from Rician noise and eddy current, leading to detail loss in reconstructing the DTI-derived parametric maps especially when sparsely sampled q-space data are used. To address this, this paper proposes a novel method, AID-DTI (\textbf{A}ccelerating h\textbf{I}gh fi\textbf{D}elity \textbf{D}iffusion \textbf{T}ensor \textbf{I}maging), to facilitate fast and accurate DTI with only six measurements. AID-DTI is equipped with a newly designed Singular Value Decomposition-based regularizer, which can effectively capture fine details while suppressing noise during network training by exploiting the correlation across DTI-derived parameters. Additionally, we introduce a Nesterov-based adaptive learning algorithm that optimizes the regularization parameter dynamically to enhance the performance. AID-DTI is an extendable framework capable of incorporating flexible network architecture. Experimental results on Human Connectome Project (HCP) data consistently demonstrate that the proposed method estimates DTI parameter maps with fine-grained details and outperforms other state-of-the-art methods both quantitatively and qualitatively.

Figures (3)

• Figure 1: The proposed AID-DTI pipeline. The network input is super sparse measurements uniformly sampled from the dense measurements, and then the mapping between the sparsely sampled signal and three DTI metrics is directly learned simultaneously. After the network output, we vectorize each parameter and concatenate them into a new matrix, then perform SVD on this matrix to obtain the singular values. The weighted parameter $\lambda$ is adaptively learned to balance between data fidelity and SVD-regularization.
• Figure 2: The ground truth, estimated DTI parameters FA, AD, and MD, and corresponding residual maps based on MF, q-DL, CNN, MESC-SD (baseline), and Ours in a test subject with 6 diffusion directions at b-values of 1000$s/{mm}^2$.
• Figure 3: Prospective results in a test subject with real low angular resolution data (6 diffusion directions at b-values of 1000$s/{mm}^2$ and 2 at $b_0$).