Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Limited-View Photoacoustic Imaging Reconstruction Via High-quality Self-supervised Neural Representation

Youshen xiao, Yuting Shen, Bowei Yao, Xiran Cai, Yuyao Zhang, Fei Gao

TL;DR

The paper tackles limited-view photoacoustic tomography (PAT) reconstruction, where restricted viewing angles cause information loss and artifacts. It proposes HIS, a high-quality self-supervised neural representation that models the target image as an implicit continuous function $I=f_\Theta(p)$ over 2D coordinates $p$ and trains by minimizing the discrepancy between predicted sensor data $\hat{y}_s = A f_\Theta$ and observed $y_s$ using Fourier feature encoding. Key contributions include applying implicit neural representations to PAT, using Fourier features to capture high-frequency content, and enabling reconstruction without external data while achieving faster performance than traditional model-based methods. Results on simulated, phantom, and in vivo data demonstrate superior reconstruction quality and robustness across limited-view scenarios, suggesting practical impact for clinical PAT where full angular coverage is impractical.

Abstract

In practical applications within the human body, it is often challenging to fully encompass the target tissue or organ, necessitating the use of limited-view arrays, which can lead to the loss of crucial information. Addressing the reconstruction of photoacoustic sensor signals in limited-view detection spaces has become a focal point of current research. In this study, we introduce a self-supervised network termed HIgh-quality Self-supervised neural representation (HIS), which tackles the inverse problem of photoacoustic imaging to reconstruct high-quality photoacoustic images from sensor data acquired under limited viewpoints. We regard the desired reconstructed photoacoustic image as an implicit continuous function in 2D image space, viewing the pixels of the image as sparse discrete samples. The HIS's objective is to learn the continuous function from limited observations by utilizing a fully connected neural network combined with Fourier feature position encoding. By simply minimizing the error between the network's predicted sensor data and the actual sensor data, HIS is trained to represent the observed continuous model. The results indicate that the proposed HIS model offers superior image reconstruction quality compared to three commonly used methods for photoacoustic image reconstruction.

Limited-View Photoacoustic Imaging Reconstruction Via High-quality Self-supervised Neural Representation

TL;DR

The paper tackles limited-view photoacoustic tomography (PAT) reconstruction, where restricted viewing angles cause information loss and artifacts. It proposes HIS, a high-quality self-supervised neural representation that models the target image as an implicit continuous function over 2D coordinates and trains by minimizing the discrepancy between predicted sensor data and observed using Fourier feature encoding. Key contributions include applying implicit neural representations to PAT, using Fourier features to capture high-frequency content, and enabling reconstruction without external data while achieving faster performance than traditional model-based methods. Results on simulated, phantom, and in vivo data demonstrate superior reconstruction quality and robustness across limited-view scenarios, suggesting practical impact for clinical PAT where full angular coverage is impractical.

Abstract

In practical applications within the human body, it is often challenging to fully encompass the target tissue or organ, necessitating the use of limited-view arrays, which can lead to the loss of crucial information. Addressing the reconstruction of photoacoustic sensor signals in limited-view detection spaces has become a focal point of current research. In this study, we introduce a self-supervised network termed HIgh-quality Self-supervised neural representation (HIS), which tackles the inverse problem of photoacoustic imaging to reconstruct high-quality photoacoustic images from sensor data acquired under limited viewpoints. We regard the desired reconstructed photoacoustic image as an implicit continuous function in 2D image space, viewing the pixels of the image as sparse discrete samples. The HIS's objective is to learn the continuous function from limited observations by utilizing a fully connected neural network combined with Fourier feature position encoding. By simply minimizing the error between the network's predicted sensor data and the actual sensor data, HIS is trained to represent the observed continuous model. The results indicate that the proposed HIS model offers superior image reconstruction quality compared to three commonly used methods for photoacoustic image reconstruction.
Paper Structure (13 sections, 6 equations, 8 figures, 1 table)

This paper contains 13 sections, 6 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Principles and methods
  3. The model-based iterative method of PAT
  4. Image Reconstruction Strategies
  5. Learning The Implicit Function
  6. Network Architecture
  7. Dataset
  8. Training Parameters
  9. Results
  10. Results of simulation data
  11. Results on phantom experiment data
  12. Results on in vivo experiment data
  13. Conclusion and discussion

Figures (8)

  • Figure 1: Commonly encountered geometric illustrations in experiments include 360° full-range tomographic imaging, along with representative angles such as 180°, 120°, 90°, and 70°.
  • Figure 2: Workflow of the proposed HIS model. The network parameterizing implicit function $f_{\Theta}$ takes the coordinate p of sampling points as input and predicts the image intensity $I=f_{\Theta}({p})$at these positions.Then,the sensor data $\bar{\hat{\mathbf{y}}}_{s}$ obtained from the predicted image are calculated by the forword operator. Finally, we optimize the network by minimizing the loss between the predicted sensor data $\bar{\hat{\mathbf{y}}}_{s}$ and real sensor data $\mathbf{y}_s$ from acquired limited-view.
  • Figure 3: The architecture of the neural network used for parameterizing the implicit function $f_{\Theta}$, which consists of the fourier encoding and a three-layers MLP.
  • Figure 4: The reconstruction results of simple geometric phantom. (a1)-(a4) represent the results of the UBP method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (b1)-(b4) are the results applying the TR method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (c1)-(c4) show the results applying the MB method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (d1)-(d4) represent the results of the HIS method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (a5)-(d5) are the same ground truth. TR, time reversal. MB, model-based. GT, ground truth.
  • Figure 5: The reconstruction results of simplified vascular phantom. (a1)-(a4) represent the results of the UBP method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (b1)-(b4) are the results applying the TR method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (c1)-(c4) show the results applying the MB method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (d1)-(d4) represent the results of the HIS method in limited-view cases of 180°, 120°, 90° and 70°, respectively. (a5)-(d5) are the same ground truth. TR, time reversal. MB, model-based. GT, ground truth.
  • ...and 3 more figures