Investigation of photoacoustic tomography reconstruction with a limited view from linear array

Мин Ай, Jiayi Cheng, Davood Karimi, Septimiu E. Salcudean, Robert Rohling, Purang Abolmaesumi, Shuo TangJournal of Biomedical Optics 2021 · DOI

TL;DR

Photoacoustic tomography (PAT) using linear arrays, common in clinical ultrasound, suffers from limited-view artifacts. This paper proposes and validates a dual-probe linear array configuration combined with an iterative VR-SGD reconstruction algorithm to significantly improve image quality and overcome these limitations.

What Problem Does This Paper Address?

Photoacoustic tomography (PAT) offers unique capabilities for biomedical imaging, particularly for visualizing blood vessels and tissue contrast. Linear array transducers, widely used in clinical ultrasound due to their availability and cost-effectiveness, are also attractive for PAT. However, a fundamental challenge with linear arrays is the "limited-view problem." Because the transducer elements are all on one side of the imaging region, they can only capture signals originating from a limited range of angles. This incomplete data leads to reconstruction artifacts, blurring, and a loss of structural detail, hindering the broader clinical application of PAT with linear arrays. Existing solutions often involve hardware modifications that may not be practical for clinical use.

Methods

The authors investigate two main strategies to address the limited-view problem:

  1. Detection Geometry Optimization: They compare several linear array configurations:
  2. Single Array: The traditional setup with one linear array.
  3. Dual-Probe Configurations: Two linear arrays positioned at an angle (specifically, parallel-shape with $\theta = 180^\circ$ and L-shape with $\theta = 90^\circ$). This aims to increase the effective detection angle.
  4. Square-Shape: Four linear arrays enclosing the object, mimicking a full-view scenario for comparison.
  5. Iterative Reconstruction Algorithm: They implement and evaluate a Variance-Reduced Stochastic Gradient Descent (VR-SGD) algorithm for image reconstruction. This is an iterative, model-based approach that aims to minimize a cost function relating the measured data to a projected image. Unlike traditional Delay-and-Sum (DAS) methods, iterative algorithms can leverage prior knowledge (the forward model) and regularization to reduce artifacts and improve image quality, especially in underdetermined systems like limited-view PAT. VR-SGD is chosen for its balance of computational efficiency and convergence speed compared to other gradient descent methods.

Datasets

The study utilizes both simulated and experimental data:

  • Simulated Data:
    • A tree branch phantom was used to mimic blood vessels.
  • A FORBILD phantom was used to provide more distinguishable features.
  • These phantoms were simulated with various detection geometries (single-array, parallel-shape, L-shape, square-shape).
  • Experimental Data:
    • Printed dots array phantom: Used to evaluate single-array detection.
  • Pencil leads phantom: Another test object for single-array detection.
  • Octagon phantom with dots: Used to compare single-array and L-shape dual-probe detection.
  • Tube phantom filled with methylene blue: Used to evaluate single-array and L-shape dual-probe detection.
  • The experimental data was acquired using a custom PAT system with linear array transducers.

Key Insights and Ideas

  • Dual-Probe is Key: The most significant finding is that a dual-probe linear array configuration, particularly the L-shape ($\theta = 90^\circ$), offers a substantial improvement in addressing the limited-view problem. It captures more complete structures and balances the trade-off between increased view angle and potential background noise.
  • Iterative Reconstruction is Crucial: The VR-SGD algorithm consistently outperforms traditional DAS reconstruction across all tested geometries, significantly reducing artifacts and noise, and improving contrast and signal-to-noise ratios.
  • L-shape + VR-SGD Synergy: The combination of the L-shape dual-probe geometry and the VR-SGD algorithm yields results that are comparable to the full-view square-shape geometry in simulations, demonstrating its potential for clinical applications where full-view is not feasible.
  • Trade-off Between Speed and Quality: While VR-SGD significantly improves image quality, it comes at the cost of increased processing time compared to the faster DAS method.

Key Results

Simulations: * For the tree branch phantom, VR-SGD reduced RMSE from 0.472 (DAS, single-array) to 0.351, and from 0.437 (DAS, square-shape) to 0.184. The L-shape dual-probe with VR-SGD achieved an RMSE of 0.258, comparable to the full-view square-shape. * For the FORBILD phantom, VR-SGD reduced RMSE from 0.609 (DAS, single-array) to 0.519, and from 0.423 (DAS, square-shape) to 0.201. Again, L-shape with VR-SGD (RMSE 0.367) was close to the square-shape. * The L-shape geometry showed comparable performance to the square-shape (full-view) in terms of RMSE and convergence rate.

Experiments: * Dots array: gCNR improved from 0.766 (DAS, single-array) to 0.956 (VR-SGD, single-array). * Pencil leads: gCNR improved from 0.644 (DAS, single-array) to 0.897 (VR-SGD, single-array). * Octagon phantom: L-shape dual-probe with VR-SGD achieved a gCNR of 0.931, compared to 0.902 for DAS. * Tube phantom: L-shape dual-probe with VR-SGD achieved a gCNR of 0.977, compared to 0.783 for DAS. * Quantitative comparisons (CNR, SNR, gCNR) in Table 2 consistently show significant improvements with VR-SGD over DAS, especially for dual-probe configurations.

Limitations and Future Work

Limitations: * Computational Cost: The VR-SGD algorithm is computationally intensive, taking hours for reconstruction compared to minutes for DAS. The authors suggest hardware upgrades (e.g., GPUs) and software optimization for future speed improvements. * Forward Model Simplification: The current forward projection model is a simplified representation. Future work could incorporate more factors like light attenuation and limited transducer bandwidth for a more accurate model. * gCNR Sensitivity: While useful, gCNR can be sensitive to ROI selection and background noise, and doesn't directly measure similarity to the ground truth.

Future Work: * Investigate other iterative reconstruction methods (e.g., TwIST, ADMM) and different regularizers. * Develop more comprehensive forward projection models, including 3D and accounting for physical phenomena like light attenuation and limited bandwidth. * Explore deep learning-based methods for PAT reconstruction, potentially using experimental data for training to handle more realistic scenarios. * Further explore the clinical applicability of dual-probe configurations in specific areas like breast and prostate imaging, considering anatomical constraints and accessibility.

Summary updated

Abstract

PAT with dual-probe detection and iterative reconstruction is a promising solution to the limited-view problem of linear arrays.
BibTeX
@article{2021investigation,
  author = {Мин Ай and Jiayi Cheng and Davood Karimi and Septimiu E. Salcudean and Robert Rohling and Purang Abolmaesumi and Shuo Tang},
  title = {{Investigation of photoacoustic tomography reconstruction with a limited view from linear array}},
  journal = {Journal of Biomedical Optics},
  year = {2021},
  doi = {10.1117/1.jbo.26.9.096009},
}
Figure