Learning-based sound speed estimation and aberration correction in linear-array photoacoustic imaging

TL;DR

This work tackles aberration artefacts in photoacoustic imaging caused by spatially varying SoS in heterogeneous tissue by introducing a learning-based framework that estimates SoS distributions from co-registered US RF data and uses them to correct PA reconstructions. A CNN encoder–decoder is pre-trained on 2D US simulations and refined via transfer learning on physical phantoms, enabling fast, accurate SoS maps that feed into a time-reversal PA reconstruction in a dual PA/US system. Quantitative results show RMSEs of $10.2$ m/s (digital) and $15.2$ m/s (physical phantoms) for SoS estimation, SSIM up to $0.86$ for PA reconstructions, and up to $1.2 imes$ SNR improvement in vivo, indicating strong potential for real-time clinical translation. The approach leverages existing clinical US probes and co-registered data to deliver robust aberration correction without heavy hardware or iterative optimization, pushing forward the practical adoption of PA imaging in preclinical and clinical workflows.

Abstract

Photoacoustic (PA) image reconstruction involves acoustic inversion that necessitates the specification of the speed of sound (SoS) within the medium of propagation. Due to the lack of information on the spatial distribution of the SoS within heterogeneous soft tissue, a homogeneous SoS distribution (such as 1540 m/s) is typically assumed in PA image reconstruction, similar to that of ultrasound (US) imaging. Failure to compensate the SoS variations leads to aberration artefacts, deteriorating the image quality. Various methods have been proposed to address this issue, but they usually involve complex hardware and/or time-consuming algorithms, hindering clinical translation. In this work, we introduce a deep learning framework for SoS estimation and subsequent aberration correction in a dual-modal PA/US imaging system exploiting a clinical US probe. As the acquired PA and US images were inherently co-registered, the estimated SoS distribution from US channel data using a deep neural network was incorporated for accurate PA image reconstruction. The framework comprised an initial pre-training stage based on digital phantoms, which was further enhanced through transfer learning using physical phantom data and associated SoS maps obtained from measurements. This framework achieved a root mean square error of 10.2 m/s and 15.2 m/s for SoS estimation on digital and physical phantoms, respectively and structural similarity index measures of up to 0.86 for PA reconstructions as compared to the conventional approach of 0.69. A maximum of 1.2 times improvement in signal-to-noise ratio of PA images was further demonstrated with a human volunteer study. Our results show that the proposed framework could be valuable in various clinical and preclinical applications to enhance PA image reconstruction.

Figures (5)

• Figure 1: Learning-based sound speed estimation and aberration correction framework for dual-modal photoacoustic (PA) /ultrasound (US) imaging. Top branch: Deep learning based speed of sound (SoS) estimation using US channel data. Bottom branch: SoS compensation for PA image reconstruction (PA images were acquired from cross sections of human fingers in vivo).
• Figure 2: Learning-based sound speed estimation and aberration correction in dual-modal photoacoustic (PA)/ultrasound (US) imaging demonstrated with digital phantoms with point optical absorbers. (a) US reconstruction. The top 2 mm are zeroed out due to the noise interference. (b) The ground truth (GT) of the Sound of Speed (SoS) and (c) SoS distributions predicted by the deep learning (DL) model. (d-e) PA reconstruction using an optimal SoS value given by an autofocus method and SoS distributions by the DL model. (f) Initial pressure distribution used for acoustic forwarding in the k-Wave simulation and PA reconstruction using the GT SoS. The reconstructed optical point sources at two different depths in PA are enlarged in the insets (d-f). Images share the same scale bar of 5 mm.
• Figure 3: Learning-based sound speed estimation and aberration correction in dual-modal photoacoustic (PA)/ultrasound (US) imaging demonstrated with agar-based tissue-mimicking phantom with pencil leads as inclusions (indicated by yellow arrows in US B-mode reconstruction). (a) US B-mode image. (b-c) SoS distributions predicted by the deep learning (DL) model. (d) Initial pressure distribution. (e-h) PA reconstructions using the SoS of soft tissues (c = 1540 m/s) and water (c = 1490 m/s), and the SoS distributions by the DL model. The reconstructed pencil leads (cross-sections) in PA are enlarged in the insets. Scale bar: 5 mm.
• Figure 4: Learning-based sound speed esitmation and aberration correction in dual-modal photoacoustic (PA)/ultrasound (US) imaging demonstrated with needle in-plane and out-of-plane insertions into porcine tissues ex vivo. (a) US B-mode image. (b-c) SoS distributions predicted by the deep learning (DL) model. (d-e) PA reconstructions using a conventional SoS value in soft tissue (c = 1540 m/s) and the SoS distributions by the DL model; Yellow arrows denote image artefacts associated with acoustic reflection and limited view. Scale bar: 5 mm.
• Figure 5: Learning-based sound speed estimation and aberration correction in dual-modal photoacoustic (PA)/ultrasound (US) imaging demonstrated with in vivo 3D volume of the human wrist. (a) US B-mode image. (b) SoS distribution by the deep learning (DL) model after transfer training (c-d) PA reconstructions using a uniform SoS value of water (c = 1500 m/s at 23 $^\circ$C) and the SoS distribution by the DL model. (e-f) 3D volumes and maximum intensity projections from X-Y and Z-Y directions. White triangular denotes the cross-section of a blood vessel. White arrow denotes vessels at deep depths. Scale bar: 5 mm.