Deep learning-based free-water correction for single-shell diffusion MRI

Tianyuan Yao, Derek B. Archer, Praitayini Kanakaraj, Nancy R. Newlin, Shunxing Bao, Daniel Moyer, Kurt G. Schilling, Bennett A. Landman, Yuankai HuoMagnetic 2025 · DOI

TL;DR

This paper introduces a novel deep learning framework for "free-water elimination" (FWE) in diffusion MRI (dMRI), a crucial step for accurate brain microstructure analysis. Their method excels at correcting for partial volume effects caused by free water, particularly in single-shell dMRI data which is common in clinical settings, outperforming existing approaches.

What Problem Does This Paper Address?

Diffusion MRI (dMRI) is a powerful tool for mapping brain connectivity and microstructure. However, a significant challenge arises from "partial volume effects" (PVE), where a single voxel in the MRI scan contains a mixture of different tissue types. A common culprit is "free water," such as cerebrospinal fluid (CSF) in the ventricles or edema in pathological conditions. This free water has different diffusion properties than brain tissue, and when not accounted for, it contaminates the dMRI signal, leading to biased estimates of key diffusion metrics like Fractional Anisotropy (FA) and Mean Diffusivity (MD). These biases can obscure subtle changes in white matter integrity, making it difficult to diagnose or track neurodegenerative diseases or brain tumors.

Traditional methods for free-water elimination (FWE) often rely on multi-shell dMRI data (acquisitions with multiple b-values), which are more robust but also more time-consuming and less common in clinical practice. Single-shell acquisitions are prevalent due to time constraints, but FWE in this scenario is mathematically ill-posed, leading to unreliable results. This paper aims to bridge this gap by developing a deep learning approach that can accurately perform FWE even with single-shell dMRI data.

Methods

The core of this work is a novel deep learning framework designed for FWE. Key components include:

  • Dynamic Head: This module allows the network to adapt to different input configurations, specifically handling varying combinations of b-values from multi-shell data. It uses a binary code to indicate the presence or absence of specific shells and dynamically generates parameters for the first convolutional layer to mitigate input variations.
  • Spherical Convolution: This technique is employed to handle the complex, non-Euclidean geometry of diffusion MRI data and achieve rotational equivariance. The architecture uses an S² convolutional layer followed by an SO(3) convolutional layer, processing directional information on the sphere and within the rotation group, respectively.
  • B-value Mapping: To harmonize data from different acquisition schemes, a linear decay model is used to re-map b-values to a consistent training scheme (e.g., 1000 s/mm²).
  • Deep Learning Architecture: The framework combines the dynamic head and spherical convolution with efficient generalized layers to estimate the free water fraction ($f_{fw}$).

This approach differs from previous deep learning methods by being designed for generalizability across various acquisition schemes, not just specific single-shell or multi-shell configurations.

Datasets

The study utilized several datasets for training and evaluation:

  • Human Connectome Project Young Adults (HCP-ya): This was the primary dataset for training and testing the model. It includes multi-shell dMRI data with b-values of 1000, 2000, and 3000 s/mm². A subset was used for training, validation, and testing. A test-retest group from HCP-ya was also used to assess reproducibility.
  • HCP Aging (HCP-a): This dataset was used to evaluate the model's generalizability to an older population and different acquisition parameters (two shells: 1500 and 3000 s/mm²).
  • Brain Tumor Connectomics Data (BTC): This dataset, including pre-operative and post-operative data from patients with brain tumors, was used to assess the method's performance in a clinical context.

The "silver standard" for target free water fractions was generated using the full multi-shell configuration of the FWE model from these datasets.

Key Insights and Ideas

  • Generalizability is Key: The paper highlights that many existing deep learning methods for FWE are not generalizable to new acquisition schemes. Their "dynamic head" approach is a significant step towards a unified model that can handle diverse inputs.
  • Single-Shell FWE is Achievable: Despite the mathematical ill-posedness of FWE in single-shell data, the deep learning approach demonstrates that accurate estimation is possible, opening doors for wider clinical adoption.
  • Free Water Correction Improves Tractography: Visualizations show that FWE can lead to more complete and accurate tractography, especially in tumor regions, by removing artifacts caused by free water.
  • Reproducibility Boost: The FWE method significantly improves the test-retest reproducibility of DTI metrics, indicating more stable and reliable measurements.

Key Results

  • Superior RMSE: The proposed deep learning method achieved the lowest RMSE for free water fraction ($f_{fw}$) estimation (1.97 ± 0.21 x 10⁻²) compared to conventional methods (3.62 ± 0.44 x 10⁻²) and other deep learning approaches like ANN (3.08 ± 0.31 x 10⁻²) and Deep SHORE (2.89 ± 0.19 x 10⁻²) on single-shell data.
  • Consistent Performance Across Shells: The proposed method showed consistent improvements across single, two, and three-shell configurations, with tighter variance, indicating higher precision.
  • Improved FA and MD: The method demonstrated comparable or improved FA values in CSF-confined ventricles compared to baselines and the silver standard. It also produced MD maps with distributions closest to the silver standard, avoiding over-regularization issues seen in other methods.
  • Enhanced Tractography: In tumor patients, FWE resulted in a notable increase in the number of streamlines detected around the tumor region (e.g., 4.2% increase in one case), suggesting better delineation of affected white matter tracts.
  • Reproducibility Gains: FWE significantly improved test-retest reproducibility of FA and MD for key white matter tracts like the corpus callosum, corticospinal tract, and superior longitudinal fascicle.

Limitations and Future Work

  • Dependence on Training Data Shell Configuration: The proposed method's generalizability is still somewhat dependent on the shell configurations present in the training data.
  • High b-value SNR Concerns: While high b-values (e.g., 3000 s/mm²) are beneficial for attenuating free water signal, they can lead to reduced SNR, potentially impacting performance on datasets with lower quality. Further exploration in this area is warranted.
  • Clinical Applicability of Shells: The study used higher b-values (up to 3000 s/mm²) for research purposes. For clinical applications, lower b-values (500-1000 s/mm²) are more common, and the framework's performance with these shells needs further investigation.
  • Other Acquisition Parameters: Differences in spatial resolution, echo time (TE), and SNR can affect generalizability. Future work should explore adaptive learning strategies or data harmonization techniques to address these variations.
  • Advanced Modeling: The linear decay model for b-value mapping is a simplification. More advanced radial space modeling could potentially improve calibration, especially for complex tissues.
  • Adaptive Head Design: Introducing an adaptive head design instead of binary coding for shell presence/absence could further enhance the model.
  • Weighted Loss Functions: Incorporating weighted loss functions could help the model better handle tissue-specific variations and regions with higher free water contamination.

Summary updated

Abstract

Free-water elimination (FWE) modeling in diffusion magnetic resonance imaging (dMRI) is crucial for accurate estimation of diffusion properties by mitigating the partial volume effects caused by free water, particularly at the interface between white matter and cerebrospinal fluid. The presence of free water partial volume effects leads to biases in estimating diffusion properties. Additionally, the existing mathematical FWE model is a two-compartment model, which can be well posed for multi-shell data. However, single-shell acquisitions are more common in clinical cohorts due to time constraints. To overcome these problems, we proposed a deep-learning framework that focuses on mapping and correcting free-water partial volume contamination in DWI. It utilizes data-driven techniques to infer plausible free-water volumes across different diffusion MRI acquisition schemes, including single-shell acquisitions. In this work, we study the Human Connectome Project Young Adults (HCP-ya), the HCP Aging dataset (HCP-a) as well as Brain Tumor Connectomics Data (BTC). The evaluation demonstrates that it produces more plausible results compared to previous single-shell free water estimation approaches. The proposed method is generalizable through model fine-tuning and b-value re-mapping when dealing with new data. The results have demonstrated improved consistency of properties estimation between scan/rescan data and accuracy in identifying neural pathways, as well as enhanced clarity in the visualization of white matter tracts.
BibTeX
@inproceedings{yao2025deep,
  author = {Tianyuan Yao and Derek B. Archer and Praitayini Kanakaraj and Nancy R. Newlin and Shunxing Bao and Daniel Moyer and Kurt G. Schilling and Bennett A. Landman and Yuankai Huo},
  title = {{Deep learning-based free-water correction for single-shell diffusion MRI}},
  booktitle = {Magnetic},
  year = {2025},
  doi = {10.1016/j.mri.2025.110326},
}
Figure