Abstract
Neurite Orientation Dispersion and Density Imaging (NODDI) is a critical technique for mapping brain microstructure, yet deep learning-based estimation methods often suffer from "direction sensitivity," where performance degrades when testing diffusion directions differ from training. We present RobNODDI, a robust framework that utilizes continuous spherical harmonic (SH) representations and adaptive sampling to overcome this limitation. By shifting the learning task from discrete signal values to a continuous mathematical space, RobNODDI ensures high accuracy and generalization across diverse clinical scanning protocols without the need for retraining.
Introduction
While deep learning (DL) has significantly accelerated NODDI parameter estimation compared to traditional optimization methods, it remains brittle when faced with variations in diffusion gradient directions. Most DL models are trained on a fixed set of directions; if a clinician uses a model on data acquired with even slightly different directions—a common occurrence across different scanner manufacturers—the model’s accuracy plummets. This study verifies this gap, demonstrating that mainstream models like q-DL and GCNN suffer significant performance drops when testing directions do not perfectly match the training set.
Methodology
RobNODDI addresses the direction sensitivity problem through three core innovations:
- Continuous Representation via Spherical Harmonics (SH): Instead of processing raw dMRI signal intensities, RobNODDI fits signals to SH coefficients. This transforms the data into a continuous representation on a sphere, making the model agnostic to the specific coordinate grid of the original scan.
- Adaptive Sampling: During training, the sampling pattern varies across different epochs. This forces the network to learn the underlying tissue physics rather than memorizing a specific sampling pattern, facilitating true domain generalization.
- Architecture: The framework employs a Sparse Dictionary-based Long Short-Term Memory (SD-LSTM) unit followed by fully connected layers, optimized to unfold iterative reconstruction processes into a neural network for multi-shell dMRI data.
Experimental Results
The study utilized the Human Connectome Project (HCP) dataset, with 100 adult subjects and ground truth generated via the AMICO framework. RobNODDI was compared against several mainstream methods under two conditions: Same Sampling (SS) and Random Sampling (RS).
Quantitative Performance
In RS tests where directions were altered, RobNODDI maintained a stable Mean Squared Error (MSE) of 0.00211. In contrast, competing models like GCNN saw their error increase nearly sixfold to 0.01268. As shown in the table below, RobNODDI consistently achieved the highest PSNR and SSIM across both sampling conditions.
Robustness and Flexibility
RobNODDI demonstrates remarkable flexibility, handling a different number of directions at test time than it encountered during training. For instance, a model trained on 30 directions can accurately estimate parameters from a 20-direction or 40-direction scan. Qualitative analysis further confirms that while traditional models produce noisy and anatomically incorrect maps for Intracellular Volume Fraction ($V_{ic}$) and Orientation Dispersion (OD) when directions are altered, RobNODDI’s maps remain clean and consistent with the gold standard.
Conclusion
RobNODDI solves the long-standing direction-consistency problem in dMRI parameter estimation. By breaking the "grid lock" of fixed training gradients through continuous representation, it provides a robust, clinical-ready solution that maintains state-of-the-art accuracy. Future work will focus on improving robustness to varying b-values and optimizing the computational overhead of the SH fitting step.
Summary updated