Abstract
Accurate mapping of brain microstructure in diffusion MRI (dMRI) is often hindered by the use of fixed response functions, which fail to account for biological diversity and tissue interfaces. This paper introduces a robust framework utilizing Response Function Groups (RFGs) and direct cardinality ($l_0$-norm) penalization. By solving a group-sparsity problem via the Non-monotone Iterative Hard Thresholding (NIHT) algorithm, the proposed method achieves more accurate fiber orientation estimates and cleaner tissue volume fractions, particularly at the interfaces between white matter, gray matter, and cerebrospinal fluid.
Introduction
Constrained Spherical Deconvolution (CSD) is a standard technique in dMRI for estimating nerve fiber orientations. However, traditional CSD typically assumes a single, spatially invariant "Response Function" (RF) for each tissue type—essentially a template of how a single fiber bundle should appear. This assumption often leads to "spurious peaks" (phantom fiber directions) when the fixed template mismatches the actual heterogeneous brain tissue. While multi-tissue CSD methods exist, they still rely on fixed models that fail to capture the natural variability of axial and radial diffusivities across different brain regions.
Methods
The authors propose a two-fold advancement to address these limitations:
Response Function Groups (RFGs)
Instead of a single template, the method employs groups of templates for each tissue type. The White Matter (WM) RFG consists of a collection of tensors with varying axial and radial diffusivities to represent different fiber densities and thicknesses. Similarly, Gray Matter (GM) and Cerebrospinal Fluid (CSF) RFGs include isotropic templates with a range of diffusivities.
Cardinality-Penalized Optimization
To select the optimal templates for a given voxel, the authors solve a least-squares problem with a cardinality ($l_0$-norm) penalty. This encourages "group sparsity," ensuring the signal is explained by the minimum necessary number of tissue groups, effectively accounting for the partial volume effect where a single voxel contains multiple tissue types.
The NIHT Algorithm
Rather than using $l_1$ (LASSO) approximations, which can be suboptimal for true sparsity and require multiple re-weighting steps, the authors utilize the Non-monotone Iterative Hard Thresholding (NIHT) algorithm. This allows for direct minimization of the $l_0$ penalty, offering higher computational efficiency and producing cleaner results.
Results
The method was validated using high-resolution multi-shell data from the Human Connectome Project (HCP).
Tissue Volume Fractions and Residuals
The proposed L200-RFG method produced tissue maps (WM, GM, CSF) that align significantly better with anatomical T1-weighted reference images than standard methods. Furthermore, fitting residuals—the error left over after the model fit—were lower and more uniform across the brain, indicating a more accurate model fit.
Improved Fiber Orientation Distribution Functions (FODFs)
A key advantage of the RFG approach is its performance at tissue interfaces. At the WM-GM boundary, the method successfully extracts directional information that is typically obscured by noise or partial volume effects in traditional models. The resulting FODFs "penetrate" more accurately into the gray matter cortex with fewer false-positive peaks.
Discussion and Conclusion
By providing a "menu" of response functions, the RFG framework adapts to local tissue variations without requiring manual tuning for different brain regions. The direct $l_0$ minimization via NIHT proves to be both faster and more biologically faithful than $l_1$ alternatives. While the model complexity increases with the use of groups, it provides a robust foundation for future work, such as incorporating axonal diameter estimation or applying the framework to pathological states like edema, where standard fixed-RF models typically fail.
Summary updated