Deep Learning for <scp>fODF</scp> Estimation in Infant Brains: Model Comparison, Ground‐Truth Impact, and Domain Shift Mitigation

R. P. Lin, Hamza Kebiri, Ali Gholipour, Yufei Chen, Jean‐Philippe Thiran, Davood Karimi, Meritxell Bach CuadraHuman 2025 · DOI

TL;DR

This paper systematically evaluates how well deep learning models predict brain white matter fiber orientations (fODFs) when faced with data from different scanners, ages, or protocols (out-of-domain or OOD). They found that U-Net excels with limited data, while fine-tuning is a powerful strategy to adapt models to new data, especially for complex fiber crossings.

What Problem Does This Paper Address?

Understanding how the brain's white matter develops is crucial, and diffusion MRI (dMRI) is a key tool for this. Estimating the direction of white matter fibers, represented by fiber orientation distribution functions (fODFs), is essential for this understanding. While deep learning (DL) models show promise for fODF estimation, their performance when applied to data that differs from their training data (out-of-domain or OOD) is largely unknown. This is a significant problem because real-world clinical and research data often comes from different scanners, acquisition protocols, or subjects of varying ages, leading to "domain shifts." This study aims to thoroughly assess the robustness of different DL architectures to these domain shifts and evaluate strategies to mitigate them.

Methods

The researchers compared three state-of-the-art DL architectures:

  • Multilayer Perceptron (MLP): A voxel-wise approach that processes each voxel independently.
  • CNN-Transformer (CTtrack): A hybrid model combining Convolutional Neural Networks (CNNs) for local feature extraction and Transformers for global context modeling.
  • U-Net/CNN: A U-Net-based architecture known for its ability to capture multiscale spatial context, particularly effective with limited data.

They used two methods to generate "ground truth" fODFs for training:

  • MSMT-CSD (Multi-Shell Multi-Tissue Constrained Spherical Deconvolution): A standard method requiring dense, multi-shell dMRI data.
  • SS3T-CSD (Single-Shell Three-Tissue Constrained Spherical Deconvolution): A more data-efficient method suitable for developing brains.

To address domain shifts, they evaluated two strategies:

  • Method of Moments (MoM): A data harmonization technique that linearly transforms target domain data to match source domain statistics.
  • Fine-tuning: A transfer learning approach where a pre-trained model is further trained on target domain data.

Datasets

The study utilized data from two large infant dMRI datasets:

  • Developing Human Connectome Project (dHCP): Provided data from neonates.
  • Baby Connectome Project (BCP): Provided data from older infants and young children.

In total, 488 subjects were used. The researchers also performed a systematic validation of their reference CSD-based reconstruction methods by splitting each subject's data into two subsets.

Key Insights and Ideas

  • U-Net's Data Efficiency: U-Net consistently outperformed other models when fewer diffusion gradient directions were available, suggesting its strong ability to leverage spatial context even with limited input data. This is crucial for infant imaging where scan times are often restricted.
  • Ground Truth Matters: The choice of ground truth reconstruction method (SS3T-CSD vs. MSMT-CSD) significantly impacted model performance, with SS3T-CSD showing better robustness to age-related domain shifts and superior performance in capturing crossing fibers.
  • Diminishing Returns: Increasing the number of input diffusion directions beyond 28-45 showed diminishing returns in accuracy for all models, indicating potential for reducing scan times without substantial performance loss.
  • Age-Related Asymmetry: Domain shifts due to age were asymmetric. Transferring models from older to younger subjects (late to early) caused more performance degradation than the reverse (early to late), highlighting the rapid changes in brain microstructure during early development.
  • Fine-tuning is Powerful: Both MoM and fine-tuning significantly improved OOD performance, but fine-tuning consistently yielded superior results, especially for the U-Net architecture. This suggests that adapting model parameters to the target domain is highly effective.

Key Results

  • Intra-Site Performance: U-Net showed superior performance with fewer diffusion directions (6 directions) across both dHCP and BCP datasets, particularly with SS3T-CSD ground truth. For higher numbers of directions (28-45), MLP and CTtrack showed marginal improvements.
  • Age-Related Domain Shift: Asymmetric age-related domain shifts were observed, with the "early to late" transfer showing more degradation. SS3T-CSD demonstrated greater robustness to these shifts compared to MSMT-CSD.
  • Inter-Site Domain Adaptation: Both MoM and fine-tuning significantly improved performance (p < 0.05) in over 95% of tested configurations for inter-site transfers (dHCP to BCP and vice versa). Fine-tuning consistently outperformed MoM, and U-Net benefited the most from increased target subjects.
  • Quantitative Improvements: Fine-tuning improved ACC (angular correlation coefficient) values to above 0.80, and significantly reduced GFA (generalized fractional anisotropy) differences, especially for the challenging BCP to dHCP transfer.

Limitations and Future Work

  • Ground Truth Limitations: The authors acknowledge that CSD-based reconstructions are not "absolute ground truth" and are themselves approximations. This inherent limitation of the reference data can bound the performance of even the best DL models.
  • Unsupervised vs. Supervised: Direct comparison with unsupervised methods is noted as methodologically problematic due to their different learning objectives.
  • Future Directions:
    • Merging reference reconstruction algorithms to leverage their strengths.
  • Extending domain adaptation to cross-shell scenarios (different b-values).
  • Optimizing the selection of target subjects for fine-tuning and exploring synthetic data generation.
  • Conducting controlled studies to disentangle biological developmental changes from technical domain shifts.
  • Further refining domain adaptation techniques, particularly in selecting optimal target subjects and improving ground-truth models.

Summary updated

Abstract

The accurate estimation of fiber orientation distribution functions (fODFs) in diffusion magnetic resonance imaging (MRI) is crucial for understanding early brain development and its potential disruptions. Although supervised deep learning (DL) models have shown promise in fODF estimation from neonatal diffusion MRI (dMRI) data, the out-of-domain (OOD) performance of these models remains largely unexplored, especially under diverse domain shift scenarios. This study evaluated the robustness of three state-of-the-art DL architectures: multilayer perceptron (MLP), transformer, and U-Net/convolutional neural network (CNN) on fODF predictions derived from dMRI data. Using 488 subjects from the developing Human Connectome Project (dHCP) and the Baby Connectome Project (BCP) datasets, we reconstructed reference fODFs from the full dMRI series using single-shell three-tissue constrained spherical deconvolution (SS3T-CSD) and multi-shell multi-tissue CSD (MSMT-CSD) to generate reference fODF reconstructions for model training, and systematically assessed the impact of age, scanner/protocol differences, and input dimensionality on model performance. Our findings reveal that U-Net consistently outperformed other models when fewer diffusion gradient directions were used, particularly with the SS3T-CSD-derived ground truth, which showed superior performance in capturing crossing fibers. However, as the number of input diffusion gradient directions increased, MLP and the transformer-based model exhibited steady gains in accuracy. Nevertheless, performance nearly plateaued from 28 to 45 input directions in all models. Age-related domain shifts showed asymmetric patterns, being less pronounced in late developmental stages (late neonates, and babies), with SS3T-CSD demonstrating greater robustness to variability compared to MSMT-CSD. To address inter-site domain shifts, we implemented two adaptation strategies: the Method of Moments (MoM) and fine-tuning. Both strategies achieved significant improvements ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:semantics><mml:mrow><mml:mi>p</mml:mi> <mml:mo><</mml:mo> <mml:mn>0.05</mml:mn></mml:mrow> <mml:annotation>$$ p<0.05 $$</mml:annotation></mml:semantics> </mml:math> ) in over 95% of tested configurations, with fine-tuning consistently yielding superior results and U-Net benefiting the most from increased target subjects. This study represents the first systematic evaluation of OOD settings in DL applications to fODF estimation, providing critical insights into model robustness and adaptation strategies for diverse clinical and research applications.
BibTeX
@inproceedings{lin2025deep,
  author = {R. P. Lin and Hamza Kebiri and Ali Gholipour and Yufei Chen and Jean‐Philippe Thiran and Davood Karimi and Meritxell Bach Cuadra},
  title = {{Deep Learning for <scp>fODF</scp> Estimation in Infant Brains: Model Comparison, Ground‐Truth Impact, and Domain Shift Mitigation}},
  booktitle = {Human},
  year = {2025},
  doi = {10.1002/hbm.70367},
}
Figure