A single-view-based electroacoustic tomography imaging using deep learning for electroporation monitoring

Abstract

Background

Electroacoustic tomography (EAT) is an emerging imaging technique with potential for guiding electroporation therapy. However, in clinical applications, EAT scan is typically limited to a single view acquisition, leading to severe artifacts in the reconstruction.

Purpose

This study aims to develop novel deep-learning models to substantially enhance the image quality of the single-view EAT reconstruction.

Methods

We proposed a two-stage deep learning model (EXPP-Net) to address the limited view problem in EAT image reconstruction. The two-stage model comprises extrapolation and post-processing sub-models. The extrapolation sub-model receives a single-view sinogram u_α acquired at any angle α and a rotation angle γ∈[−60°, −24°, 24°, 60°], and predicts four single-view sinograms at angles α+γ. The 5 sinograms (i.e., u_α-60°, u_α-24°, u_α, u_α+24°, u_α+60°) are back-projected to reconstruct an initial sparse-view EAT image (m_sv). The post-processing sub-model enhances m_sv’s quality to a full-view image (m_fv). These sub-models were trained using experimental data. Using a linear-array probe, 54 full-view datasets (60 or 90 views each) were acquired by delivering electrical pulses via two tungsten electrodes rotated from −180° to 180° in a water tank. Each view u_α, together with {u_α-60°, u_α-24°, u_α+24°, u_α+60°} and m_fv, forms a view set for model training. Data splitting was performed at the dataset level, yielding 34/10/10 datasets for training/validation/test, corresponding to 2190/600/600 view sets. The model performance was evaluated using root mean square error (RMSE), structural similarity index measure (SSIM), and the iso-pressure line Dice similarity coefficients (DICE). Statistical analysis used a linear mixed-effects model with Benjamini–Hochberg FDR correction (α = 0.05), and effect size was defined as standardized fixed-effect coefficients (β/σ). For comprehensive evaluation, the model was re-trained and evaluated on convex-array probe data, and further tested in vivo on one mouse. The model was also compared with a previously published method to demonstrate its advantages.

Results

The quality of EAT images reconstructed from single-view acquisition was substantially improved by EXPP-Net, showing good agreement with full-view images (linear-array probe study: RMSE: 0.0033 ± 0.0023, SSIM: 0.9968 ± 0.0059, median DICE ≥ 0.9450; convex-array probe study: RMSE: 0.0083 ± 0.0034, SSIM: 0.9958 ± 0.0038, median DICE ≥ 0.8837). In vivo results demonstrated substantial distortion correction. Comparison with the prior method on RMSE, SSIM, and DICE yielded large effect sizes (|β/σ| > 0.8), with all comparisons reaching statistical significance (p < 0.05).

Conclusions

The proposed model generated full-view EAT imaging from a single-view sinogram, substantially improving the quality and precision of EAT. These results suggest its potential for real-time electroporation verification.