Deep Learning for fODF Estimation in Infant Brains: Model Comparison, Ground-Truth Impact, and Domain Shift
Rizhong Lin1,2,3, Hamza Kebiri2,4, Ali Gholipour5,6,7
1Signal Processing Laboratory (LTS5), École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland.
Human Brain Mapping
|October 7, 2025
Summary
This study tested deep learning models for estimating brain fiber orientation from diffusion MRI data. U-Net performed best with fewer directions, while MLP and transformers improved with more data, showing adaptation strategies enhance robustness.
Area of Science:
- Neuroimaging
- Diffusion Magnetic Resonance Imaging (dMRI)
- Computational Neuroscience
Background:
- Accurate estimation of fiber orientation distribution functions (fODFs) is vital for studying brain development using dMRI.
- Supervised deep learning (DL) models show promise for fODF estimation in neonatal dMRI.
- The out-of-domain (OOD) performance and robustness of these DL models under domain shifts are not well understood.
Purpose of the Study:
- To evaluate the robustness of three DL architectures (MLP, transformer, U-Net/CNN) for fODF estimation from dMRI data.
- To assess the impact of age, scanner/protocol differences, and input dimensionality on model performance.
- To investigate adaptation strategies for improving OOD performance in DL-based fODF estimation.
Main Methods:
- Utilized dHCP and BCP datasets (488 subjects) for training and evaluation.
- Reconstructed reference fODFs using single-shell three-tissue constrained spherical deconvolution (SS3T-CSD) and multi-shell multi-tissue CSD (MSMT-CSD).
- Systematically assessed DL model performance under varying age groups, scanner protocols, and input data dimensionality. Investigated Method of Moments (MoM) and fine-tuning for domain adaptation.
Main Results:
- U-Net outperformed other models with fewer diffusion gradient directions, especially with SS3T-CSD ground truth.
- MLP and transformer models showed improved accuracy with increased input directions, with performance plateauing around 28-45 directions.
- Age-related domain shifts were less pronounced in later developmental stages; SS3T-CSD was more robust than MSMT-CSD.
- Both MoM and fine-tuning significantly improved performance across configurations (p < 0.05), with fine-tuning being superior.
Conclusions:
- This study provides the first systematic evaluation of OOD performance for DL-based fODF estimation in dMRI.
- U-Net demonstrated superior performance with limited data, while MLP and transformers scaled better with more input directions.
- Fine-tuning emerged as a highly effective strategy for mitigating inter-site domain shifts, significantly benefiting U-Net.
- Findings offer critical insights for developing robust DL models for diverse clinical and research applications in neuroimaging.
Keywords:
constrained spherical deconvolution (CSD)deep learningdiffusion MRIdomain adaptationdomain shiftfiber orientation distribution function (fODF)infantneonate
