TW-BAG: Tensor-wise Brain-aware Gate Network for Inpainting Disrupted Diffusion Tensor Imaging

Zihao Tang, Xinyi Wang, Lihaowen Zhu, Mariano Cabezas, Dongnan Liu, Michael Barnett, Weidong Cai, Chenyu Wang

Abstract

Diffusion Tensor Imaging (DTI) is a vital tool for mapping the brain's white matter, yet clinical scans are frequently compromised by motion artifacts or technical constraints, leading to missing data. We present TW-BAG (Tensor-wise Brain-aware Gate), a deep learning framework designed to reconstruct missing 3D brain structures in DTI volumes. By utilizing brain-aware gating and independent decoders for each tensor coefficient, TW-BAG achieves high-fidelity inpainting. Our results demonstrate that the reconstruction error is significantly lower than the variance associated with neurodegenerative diseases, suggesting that inpainted scans are viable for clinical research.

Introduction

Diffusion Weighted Imaging (DWI) tracks the microscopic movement of water molecules to characterize the brain's white matter architecture. This data is typically modeled as a Diffusion Tensor—a $3 \times 3$ symmetric matrix at every voxel describing the direction and magnitude of diffusion. However, real-world clinical scans are often "disrupted." Patient movement, time constraints, or technical artifacts frequently result in missing slices, particularly at the superior aspect of the head.

In many research contexts, these incomplete scans are discarded to avoid biasing results, leading to a significant loss of valuable data. Standard image inpainting techniques (e.g., content-aware fill) are often insufficient for DTI because the data is not a simple image but a complex 3D volume where each point contains multi-dimensional mathematical constraints. To address this, we propose the Tensor-Wise Brain-Aware Gate (TW-BAG) network.

Methods

The TW-BAG framework is built upon a modified U-Net architecture, incorporating three primary innovations to handle the unique requirements of neuroimaging:

1. Brain-Aware Gating (BAG)

Unlike general-purpose inpainting, TW-BAG leverages anatomical priors. The model utilizes a "brain mask" derived from standard T1-weighted structural scans, which are typically higher quality and easier to acquire. This mask guides the network to focus exclusively on valid brain tissue, ignoring the empty space outside the skull and preventing boundary artifacts.

2. Gated Convolutions

Standard convolutional filters treat all pixels equally, which is problematic when large regions of data are missing. TW-BAG employs gated convolutions that learn a dynamic mask to distinguish between valid observed pixels and the "holes" that require synthesis. This allows the model to better utilize the surrounding context to fill in missing regions.

3. Tensor-Wise (TW) Decoders

A diffusion tensor is defined by six unique coefficients ($D_{xx}, D_{xy}, D_{xz}, D_{yy}, D_{yz}, D_{zz}$). A key insight of this work is that these coefficients represent different physical properties and spatial patterns. Rather than using a single decoder for the entire tensor, TW-BAG utilizes six independent decoders—one for each coefficient. This decoupling allows the model to learn the specific spatial constraints unique to each component, resulting in a 55% improvement in Fractional Anisotropy (FA) accuracy compared to monolithic approaches.

Experimental Setup

The study utilized the Human Connectome Project (HCP) dataset, involving 100 subjects. To simulate clinical disruptions, the authors intentionally cropped the top 10% (approximately 15 slices) of the brain volumes. Data was registered to a standard MNI space and corrected for motion and eddy currents using standard neuroimaging tools (FSL).

Results and Discussion

TW-BAG demonstrated significant improvements in both quantitative accuracy and qualitative visual fidelity:

  • Accuracy: The Mean Absolute Error (MAE) of the tensor coefficients was reduced from 0.0041 in the raw cropped data to 0.0009.
  • Signal Quality: The Peak Signal-to-Noise Ratio (PSNR) improved by more than 10 dB over the baseline.
  • Efficiency: The model requires only approximately 28% of the valid brain as context to successfully reconstruct the missing 10% of superior slices.
  • Visual Fidelity: Qualitatively, the inpainted FA maps—which visualize the direction of nerve fibers—appeared sharper and more anatomically consistent than those produced by competing methods like U-VQVAE.

Crucially, the authors compared the reconstruction error to the biological variance seen in clinical populations. They found that the error introduced by TW-BAG is significantly smaller than the "gap" caused by diseases like Multiple Sclerosis or Alzheimer’s. This suggests that inpainted scans are accurate enough to be used in clinical studies without masking the underlying pathology.

Conclusion

TW-BAG provides a robust solution for recovering disrupted DTI data by combining anatomical awareness with tensor-specific learning. While this study focused on missing superior slices, the "brain-aware gate" approach holds potential for inpainting other types of brain damage, such as lesions or stroke-affected areas. Future work will evaluate the model's performance on lower-resolution clinical data and more varied types of spatial disruptions.

Summary updated

Abstract

Diffusion Weighted Imaging (DWI) is an advanced imaging technique commonly used in neuroscience and neurological clinical research through a Diffusion Tensor Imaging (DTI) model. Volumetric scalar metrics including fractional anisotropy, mean diffusivity, and axial diffusivity can be derived from the DTI model to summarise water diffusivity and other quantitative microstructural information for clinical studies. However, clinical practice constraints can lead to sub-optimal DWI acquisitions with missing slices (either due to a limited field of view or the acquisition of disrupted slices). To avoid discarding valuable subjects for group-wise studies, we propose a novel 3D Tensor-Wise Brain-Aware Gate network (TW-BAG) for inpainting disrupted DTIs. The proposed method is tailored to the problem with a dynamic gate mechanism and independent tensor-wise decoders. We evaluated the proposed method on the publicly available Human Connectome Project (HCP) dataset using common image similarity metrics derived from the predicted tensors and scalar DTI metrics. Our experimental results show that the proposed approach can reconstruct the original brain DTI volume and recover relevant clinical imaging information.
BibTeX
@misc{tang2022tw,
  author = {Zihao Tang and Xinyi Wang and Lihaowen Zhu and Mariano Cabezas and Dongnan Liu and Michael Barnett and Weidong Cai and Chenyu Wang},
  title = {{TW-BAG: Tensor-wise Brain-aware Gate Network for Inpainting Disrupted Diffusion Tensor Imaging}},
  year = {2022},
  eprint = {2210.17076},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.17076},
}
Figure