BREIT: A Framework for Brain Stroke Reconstruction using Multi-Frequency 3D EIT

Technical Summary: BREIT and dFNO-bar for 3D MF-EIT Stroke Reconstruction

Problem Statement
Multi-Frequency Electrical Impedance Tomography (MF-EIT) offers a promising, non-invasive, and portable alternative to CT/MRI for rapid stroke assessment, capable of continuous monitoring via scalp electrodes. However, the development of 3D deep-learning reconstruction methods for MF-EIT is hindered by two primary issues: (1) a lack of large-scale datasets containing paired ground-truth (GT) volumes and boundary measurements, and (2) non-standardized pipelines for data generation, simulation, and evaluation. Existing literature often relies on inconsistent modeling assumptions, simplified 2D workflows, or lacks integrated tooling for systematic benchmarking, making reproducible comparisons across different stroke types and datasets difficult.

Methodology
The authors introduce BREIT, a modular framework designed to unify radiology-based admittivity generation, realistic MF-EIT simulation, and deep learning benchmarking. The framework operates through the following pipeline:

1. Neuroimaging-to-EIT Pipeline:

   • Data Source: The pipeline is anchored on the public UCLH MF-EIT stroke cohort but extends to external MRI/CT datasets (ischemic, hemorrhagic, and healthy) to augment training data.
   • Segmentation & Mapping: Subject-specific CT/MRI scans are registered to the MNI152 standard space. Tissue segmentation (Gray Matter, White Matter, CSF) and lesion segmentation (Ischemic or Hemorrhagic) are performed using tools like FSL, ANTs, and DeepBleed.
   • Admittivity Generation: A 4-class segmentation is mapped to frequency-dependent complex admittivity ($\gamma = \sigma + i\omega\epsilon$) using literature values across 17 frequencies (5–2000 Hz), generating 17 GT admittivity volumes per subject.

2. Forward Simulation (3D CEM):

   • A self-contained Python solver implements the 3D Complete Electrode Model (CEM) using the Finite Element Method (FEM).
   • It simulates MF-EIT voltages on realistic head meshes (fine mesh $\approx$1M tetrahedra for simulation, coarse mesh $\approx$75k for reconstruction) using the UCLH electrode geometry and injection protocols.
   • The solver was validated against the EIDORS library and clinical UCLH measurements, showing high correlation ($r \approx 0.82$ mean) and low normalized mean squared error.

3. Reconstruction Algorithm (dFNO-bar):

   • 3D D-bar Foundation: The framework implements a 3D D-bar reconstruction method supporting non-uniform electrode layouts. It utilizes spherical Voronoi surface-area weights to handle non-uniform electrode distributions and computes scattering data ($t(\xi)$) using $t_{exp}$ and $t_0$ approximations.
   • dFNO-bar Architecture: The proposed dFNO-bar model integrates Fourier Neural Operators (FNO) with the D-bar structure.
     • Input: It accepts complex scattering data $t_0(\xi)$ computed at four different truncation radii ($T_\xi \in \{5, 7, 9, 11\}$), stacking real and imaginary parts to form an 8-channel input volume.
     • Processing: A 3D FNO block learns the mapping from the scattering domain to a coarse conductivity volume. This is followed by a lightweight 3D U-Net decoder that refines the prediction by concatenating the coarse output with the original 8-channel scattering input (total 9 channels).
     • Output: The final output is the reconstructed conductivity $\sigma(x) = \Re\{\gamma\}$.

Key Contributions

1. Multimodal Data Generation: A pipeline converting CT/MRI to 3D frequency-dependent admittivity volumes paired with simulated MF-EIT voltages, addressing the scarcity of GT data.
2. 3D CEM + D-bar Tooling: A Python implementation of 3D D-bar components, including support for non-uniform electrode layouts via Voronoi weighting.
3. dFNO-bar Model: A novel reconstruction model combining multi-truncation D-bar scattering data with FNO and U-Net refinement to learn the scattering-to-conductivity mapping.
4. Modular Codebase: An end-to-end, reproducible framework for generation, training, and evaluation across stroke types, facilitating standardized benchmarking.

Results
The authors evaluated dFNO-bar against classical baselines (One-step Gauss-Newton, D-bar) and a deep learning baseline (Deep D-bar) using synthetic data (250 ischemic, 219 hemorrhagic, 193 healthy cases) with added noise (50 dB, 40 dB).

• Quantitative Performance: dFNO-bar achieved the highest Structural Similarity Index (SSIM) across all noise levels. For example, in noiseless conditions, dFNO-bar achieved an SSIM of 44.13 compared to Deep D-bar's 40.09 and D-bar's 0.25. Correlation Coefficient (CC) was comparable to Deep D-bar in noisy settings but slightly higher in noiseless conditions.
• Statistical Significance: Paired tests showed significant SSIM gains for dFNO-bar over Deep D-bar ($p \ll 0.01$) across all noise settings.
• Efficiency: dFNO-bar utilized fewer parameters (67.2M vs. 90.3M) and less memory (0.59 GB vs. 0.69 GB) than Deep D-bar, though with a slightly longer inference time (0.32s vs. 0.23s).
• Visual Quality: Qualitative comparisons demonstrated that dFNO-bar provided more accurate predictions of lesion location, shape, and brain structure compared to competing methods.

Significance
The paper positions BREIT as a critical step toward reproducible and standardized research in 3D MF-EIT stroke reconstruction. By providing a unified framework that bridges neuroimaging, realistic simulation, and deep learning, it addresses the fragmentation in current modeling assumptions and simulation details. The authors claim that dFNO-bar demonstrates that integrating Fourier Neural Operators into the D-bar framework can improve reconstruction quality (specifically SSIM) while maintaining efficiency. The work aims to accelerate the development of robust 3D MF-EIT methods for time-sensitive stroke assessment, with future work planned to address clinical evaluation via partial-boundary methods and broader anatomical diversity.