Robustness of a partially learned photoacoustic reconstruction algorithm

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

6 Downloads (Pure)

Abstract

Classical non-learned algorithms for photoacoustic tomography (PAT) reconstructions are mathematically proven to converge, but they can be very slow and inadequate with respect to model and data assumptions. Recently, learned neural networks have shown to surpass the reconstruction quality of non-learned algorithms, but since analysis is challenging, convergence and stability are not guaranteed. To bridge this gap, we investigate the stability of algorithms in which we combine the structure of model-based algorithms with the efficiency of data-driven neural networks. In the last decade, primal-dual algorithms have become popular due to their ability to employ non-smooth regularisation, which is used to overcome the limited sampling problem in photoacoustic tomography. The algorithm performs updates in both the image domain (primal) and the data domain (dual). These are connected by the photoacoustic operator, which modelling is based on the laws of physics and system settings. In our approach, we replace the updates with shallow neural networks, while maintaining the primal-dual structure and the information from the photoacoustic operator. This greatly improves reconstruction quality, especially in cases of strong noise and limited sampling. This has the additional benefit that a hand-crafted regularisation does not have to be chosen, but is learned in a data-driven manner. We show its robustness in simulation and experiment against uncertainty and changes in PAT system settings. This includes the number, placement and calibration of detectors, but also changes in the tissue type that is imaged. The method is stable, computationally efficient and applicable to a generic photoacoustic system with universal applications.

Original languageEnglish
Title of host publicationPhotons Plus Ultrasound
Subtitle of host publicationImaging and Sensing 2019
EditorsLihong V. Wang, Alexander A. Oraevsky
Place of PublicationBellingham, WA
PublisherSPIE
Number of pages7
Volume10878
ISBN (Electronic)9781510623996
ISBN (Print)9781510623989
DOIs
Publication statusPublished - 6 Mar 2019
EventPhotons Plus Ultrasound: Imaging and Sensing 2019 - The Moscone Center, San Francisco, United States
Duration: 2 Feb 20197 Feb 2019

Publication series

NameProceedings of SPIE
PublisherSPIE
Volume10878
ISSN (Print)1605-7422
ISSN (Electronic)2410-9045

Conference

ConferencePhotons Plus Ultrasound: Imaging and Sensing 2019
CountryUnited States
CitySan Francisco
Period2/02/197/02/19

Fingerprint

Photoacoustic effect
Tomography
tomography
Neural networks
sampling
Sampling
operators
Physics
Calibration
Uncertainty
Noise
Hand
Tissue
Detectors
physics
detectors
simulation
Experiments

Keywords

  • Iterative image reconstruction
  • Neural networks
  • Photoacoustic tomography
  • Robustness

Cite this

Boink, Y. E., Brune, C., & Manohar, S. (2019). Robustness of a partially learned photoacoustic reconstruction algorithm. In L. V. Wang, & A. A. Oraevsky (Eds.), Photons Plus Ultrasound: Imaging and Sensing 2019 (Vol. 10878). [108781D] (Proceedings of SPIE; Vol. 10878). Bellingham, WA: SPIE. https://doi.org/10.1117/12.2507446
Boink, Yoeri E. ; Brune, Christoph ; Manohar, Srirang. / Robustness of a partially learned photoacoustic reconstruction algorithm. Photons Plus Ultrasound: Imaging and Sensing 2019. editor / Lihong V. Wang ; Alexander A. Oraevsky. Vol. 10878 Bellingham, WA : SPIE, 2019. (Proceedings of SPIE).
@inproceedings{ef0908f562c64a3a976f28189753096b,
title = "Robustness of a partially learned photoacoustic reconstruction algorithm",
abstract = "Classical non-learned algorithms for photoacoustic tomography (PAT) reconstructions are mathematically proven to converge, but they can be very slow and inadequate with respect to model and data assumptions. Recently, learned neural networks have shown to surpass the reconstruction quality of non-learned algorithms, but since analysis is challenging, convergence and stability are not guaranteed. To bridge this gap, we investigate the stability of algorithms in which we combine the structure of model-based algorithms with the efficiency of data-driven neural networks. In the last decade, primal-dual algorithms have become popular due to their ability to employ non-smooth regularisation, which is used to overcome the limited sampling problem in photoacoustic tomography. The algorithm performs updates in both the image domain (primal) and the data domain (dual). These are connected by the photoacoustic operator, which modelling is based on the laws of physics and system settings. In our approach, we replace the updates with shallow neural networks, while maintaining the primal-dual structure and the information from the photoacoustic operator. This greatly improves reconstruction quality, especially in cases of strong noise and limited sampling. This has the additional benefit that a hand-crafted regularisation does not have to be chosen, but is learned in a data-driven manner. We show its robustness in simulation and experiment against uncertainty and changes in PAT system settings. This includes the number, placement and calibration of detectors, but also changes in the tissue type that is imaged. The method is stable, computationally efficient and applicable to a generic photoacoustic system with universal applications.",
keywords = "Iterative image reconstruction, Neural networks, Photoacoustic tomography, Robustness",
author = "Boink, {Yoeri E.} and Christoph Brune and Srirang Manohar",
year = "2019",
month = "3",
day = "6",
doi = "10.1117/12.2507446",
language = "English",
isbn = "9781510623989",
volume = "10878",
series = "Proceedings of SPIE",
publisher = "SPIE",
editor = "Wang, {Lihong V.} and Oraevsky, {Alexander A.}",
booktitle = "Photons Plus Ultrasound",
address = "United States",

}

Boink, YE, Brune, C & Manohar, S 2019, Robustness of a partially learned photoacoustic reconstruction algorithm. in LV Wang & AA Oraevsky (eds), Photons Plus Ultrasound: Imaging and Sensing 2019. vol. 10878, 108781D, Proceedings of SPIE, vol. 10878, SPIE, Bellingham, WA, Photons Plus Ultrasound: Imaging and Sensing 2019, San Francisco, United States, 2/02/19. https://doi.org/10.1117/12.2507446

Robustness of a partially learned photoacoustic reconstruction algorithm. / Boink, Yoeri E.; Brune, Christoph; Manohar, Srirang.

Photons Plus Ultrasound: Imaging and Sensing 2019. ed. / Lihong V. Wang; Alexander A. Oraevsky. Vol. 10878 Bellingham, WA : SPIE, 2019. 108781D (Proceedings of SPIE; Vol. 10878).

Research output: Chapter in Book/Report/Conference proceedingConference contributionAcademicpeer-review

TY - GEN

T1 - Robustness of a partially learned photoacoustic reconstruction algorithm

AU - Boink, Yoeri E.

AU - Brune, Christoph

AU - Manohar, Srirang

PY - 2019/3/6

Y1 - 2019/3/6

N2 - Classical non-learned algorithms for photoacoustic tomography (PAT) reconstructions are mathematically proven to converge, but they can be very slow and inadequate with respect to model and data assumptions. Recently, learned neural networks have shown to surpass the reconstruction quality of non-learned algorithms, but since analysis is challenging, convergence and stability are not guaranteed. To bridge this gap, we investigate the stability of algorithms in which we combine the structure of model-based algorithms with the efficiency of data-driven neural networks. In the last decade, primal-dual algorithms have become popular due to their ability to employ non-smooth regularisation, which is used to overcome the limited sampling problem in photoacoustic tomography. The algorithm performs updates in both the image domain (primal) and the data domain (dual). These are connected by the photoacoustic operator, which modelling is based on the laws of physics and system settings. In our approach, we replace the updates with shallow neural networks, while maintaining the primal-dual structure and the information from the photoacoustic operator. This greatly improves reconstruction quality, especially in cases of strong noise and limited sampling. This has the additional benefit that a hand-crafted regularisation does not have to be chosen, but is learned in a data-driven manner. We show its robustness in simulation and experiment against uncertainty and changes in PAT system settings. This includes the number, placement and calibration of detectors, but also changes in the tissue type that is imaged. The method is stable, computationally efficient and applicable to a generic photoacoustic system with universal applications.

AB - Classical non-learned algorithms for photoacoustic tomography (PAT) reconstructions are mathematically proven to converge, but they can be very slow and inadequate with respect to model and data assumptions. Recently, learned neural networks have shown to surpass the reconstruction quality of non-learned algorithms, but since analysis is challenging, convergence and stability are not guaranteed. To bridge this gap, we investigate the stability of algorithms in which we combine the structure of model-based algorithms with the efficiency of data-driven neural networks. In the last decade, primal-dual algorithms have become popular due to their ability to employ non-smooth regularisation, which is used to overcome the limited sampling problem in photoacoustic tomography. The algorithm performs updates in both the image domain (primal) and the data domain (dual). These are connected by the photoacoustic operator, which modelling is based on the laws of physics and system settings. In our approach, we replace the updates with shallow neural networks, while maintaining the primal-dual structure and the information from the photoacoustic operator. This greatly improves reconstruction quality, especially in cases of strong noise and limited sampling. This has the additional benefit that a hand-crafted regularisation does not have to be chosen, but is learned in a data-driven manner. We show its robustness in simulation and experiment against uncertainty and changes in PAT system settings. This includes the number, placement and calibration of detectors, but also changes in the tissue type that is imaged. The method is stable, computationally efficient and applicable to a generic photoacoustic system with universal applications.

KW - Iterative image reconstruction

KW - Neural networks

KW - Photoacoustic tomography

KW - Robustness

UR - http://www.scopus.com/inward/record.url?scp=85065443122&partnerID=8YFLogxK

U2 - 10.1117/12.2507446

DO - 10.1117/12.2507446

M3 - Conference contribution

SN - 9781510623989

VL - 10878

T3 - Proceedings of SPIE

BT - Photons Plus Ultrasound

A2 - Wang, Lihong V.

A2 - Oraevsky, Alexander A.

PB - SPIE

CY - Bellingham, WA

ER -

Boink YE, Brune C, Manohar S. Robustness of a partially learned photoacoustic reconstruction algorithm. In Wang LV, Oraevsky AA, editors, Photons Plus Ultrasound: Imaging and Sensing 2019. Vol. 10878. Bellingham, WA: SPIE. 2019. 108781D. (Proceedings of SPIE). https://doi.org/10.1117/12.2507446