Microstructure Quantification by Q-Space Trajectory Imaging via Unrolled Neural Networks: Exploring Model Generalizability

Motivation: Existing machine learning models for microparameter estimation in Q-Space Trajectory Imaging (QTI) suffer from limited generalizability, particularly when the q-space encoding varies between acquisition protocols. Goal(s): To develop deep learning models that enhance computational efficiency and model generalization compared to voxel-wise training using multi-layer perceptrons. Approach: We developed a convolutional neural network by unrolling the iterative hard-thresholding algorithm and applied a novel training strategy that involves random permutation (RP) of diffusion measurements. Results: The new framework demonstrated superior efficiency and performance across evaluation metrics, with the RP strategy enabling effective learning on perturbated training data across variable q-space encoding schemes. Impact: The developed network accelerates model training with higher fidelity than state-of-the-art machine learning methods. The RP approach enhances model robustness and supports generalizability, facilitating on-the-fly QTI microstructural estimation across different acquisition protocols, which may improve clinical utility.