Author: Yankun Lang, Lei Ren, Leshan Sun, Liangzhong Xiang, Yifei Xu, Jie Zhang 👨🔬
Affiliation: University of Maryland School of Medicine, University of California, Irvine 🌍
Purpose: To achieve the full-view image from a single-view sinogram using a two-stage deep learning model for electroacoustic-tomography (EAT), which is an emerging imaging technique with significant potential for guiding electroporation therapy.
Methods: The two-stage deep learning model consists of two sub-models: an extrapolation sub-model and a post-processing sub-model. The extrapolation sub-model receives a single-view sinogram acquired at any angle (denoted as α, its sinogram denoted as uα) and a rotation angle γ (γ∈[-60°, -24°, 24°, 60°]), and then predicts four single-view sinograms at angles α+γ. The 5 sinograms (i.e., uα-60°, uα-24°, uα, uα+24° and uα+60°) are used to reconstruct initial EAT images using the backprojection algorithm, generating a sparse-view image (msv). The post-processing sub-model enhances msv to a full-view image (mfv). These sub-models were trained using the experiment data, including 54 sets of full-view ultrasound signals generated by electrical pulses delivered to the objects via two tungsten electrodes. The two electrodes were placed in a water tank and rotated from -180° to 180°. Among the 54 sets of full-view signals, uα, {uα-60°, uα-24°, uα+24°, uα+60°} and their corresponding mfv composed one sample pair. We acquired 3390 sample pairs, and split them into 2044, 1009, and 337 pairs for training, validation and testing, respectively. The model performance was evaluated using normalized root mean square error (nRMSE), structural similarity index measure (SSIM) and peak-signal-to-noise-ratio (PSNR).
Results: The image quality of the single-view EAT image was substantially improved by the two-stage model, comparable to the full-view image. The enhanced images achieved nRMSE of 0.0027±0.0019, SSIM of 0.9979±0.0051 and PSNR of 52.4333±3.8087dB.
Conclusion: The proposed model generated full-view EAT imaging from a single-view sinogram. The model is also applicable to other imaging modalities to enable single-view or limited-view image reconstruction to enhance the imaging speed with high image quality.