SamRobNODDI: Q-Space Sampling-Augmented Continuous Representation Learning for Robust and Generalized NODDI

Taohui Xiao, Jian Cheng, Wenxin Fan, Enqing Dong, Hairong Zheng, Shanshan WangarXiv 2024 · DOI · arXiv

Neurite Orientation Dispersion and Density Imaging (NODDI) is a vital MRI technique for mapping brain microstructure, providing insights into axonal density and orientation. While Deep Learning (DL) has significantly accelerated the estimation of NODDI parameters, existing models are often "brittle." They typically require fixed acquisition protocols, failing when the number of gradient directions or their coordinates change—a common occurrence in clinical settings due to protocol variations or patient motion.

This paper introduces SamRobNODDI, a sampling-agnostic framework designed to ensure robust and generalized NODDI parameter estimation across diverse diffusion MRI protocols.

The SamRobNODDI Framework

To overcome the limitations of discrete signal processing, SamRobNODDI leverages continuous mathematical representations and a novel training strategy to decouple the neural network from specific sampling patterns.

The SamRobNODDI architecture illustrating the training stage with sampling augmentation and consistency loss, and the testing stage for robust NODDI map prediction.

Key Innovations

  1. Q-Space Sampling Augmentation: During training, the model processes two versions of the same voxel data: one using standard uniform subsampling and another using random subsampling where directions vary every iteration. This exposure to diverse schemes teaches the model to handle arbitrary acquisition protocols.
  2. Continuous Representation Learning: Instead of raw signal intensities, the framework converts signals into Spherical Harmonics (SH). SH provides a continuous description of the signal on a sphere, allowing the network to learn the underlying physical "shape" of the diffusion rather than just memorizing a vector of numbers.
  3. Sampling Consistency Loss ($L_{consis}$): A specialized loss function penalizes the model if its predictions diverge between the uniform and random sampling inputs for the same voxel. This acts as a powerful regularizer, forcing the model to prioritize consistency and robustness.

Experimental Setup and Results

The framework was evaluated using data from the Human Connectome Project (HCP), utilizing 50 subjects (30 for training, 20 for testing). The ground truth was established using the AMICO algorithm on a high-density 270-direction dataset.

Performance Highlights

  • Superior Generalization: In "Random Sampling" tests—where test directions differed entirely from training—SamRobNODDI significantly outperformed state-of-the-art models like HGT and MESC-SD, which often saw performance collapse under protocol shifts.
  • Quantitative Accuracy: The model achieved the highest PSNR and SSIM scores across all three primary NODDI parameters: intracellular volume fraction ($V_{ic}$), isotropic volume fraction ($V_{iso}$), and orientation dispersion index ($OD$).
  • Clinical Efficiency: SamRobNODDI maintains high-quality results with as few as 40 total directions. This represents a 6.75x acceleration over the full 270-direction acquisition, making high-fidelity microstructure imaging feasible in time-constrained clinical environments.

Discussion and Future Work

By focusing on the underlying physical signal shape through Spherical Harmonics and enforcing sampling consistency, SamRobNODDI provides a "backbone-independent" solution. Researchers can swap the underlying neural network (e.g., from ResNet to Transformers) while retaining the robustness benefits.

While highly effective, the model's performance begins to degrade when the total number of directions drops below 30, or when individual shells have fewer than 10-15 directions. Future research will explore more complex backbone architectures and the application of this framework to other microstructural models beyond NODDI to further validate its generalized physical priors.

Summary updated

BibTeX
@misc{xiao2024samrobnoddi,
  author = {Taohui Xiao and Jian Cheng and Wenxin Fan and Enqing Dong and Hairong Zheng and Shanshan Wang},
  title = {{SamRobNODDI: Q-Space Sampling-Augmented Continuous Representation Learning for Robust and Generalized NODDI}},
  year = {2024},
  eprint = {2411.06444},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.06444},
}
Figure