[Paper Review] DALnet: High-resolution photoacoustic projection imaging using deep learning

Johannes Schwab, Stephan Antholzer+2 more|arXiv (Cornell University)|Jan 20, 2018

Photoacoustic and Ultrasonic Imaging5 references8 citations

TL;DR

This paper proposes DALnet, a deep learning-based framework that combines dynamic aperture length (DAL) correction with a convolutional neural network (CNN) to enable fast, high-resolution photoacoustic projection imaging. By addressing under-sampling, blurring, and limited-view artifacts in sparse detector arrays, DALnet produces accurate 3D images in fractions of a second without iterative reconstruction.

ABSTRACT

Photoacoustic tomography (PAT) is an emerging and non-invasive hybrid imaging modality for visualizing light absorbing structures in biological tissue. The recently invented PAT systems using arrays of 64 parallel integrating line detectors allow capturing photoacoustic projection images in fractions of a second. Standard image formation algorithms for this type of setup suffer from under-sampling due to the sparse detector array, blurring due to the finite impulse response of the detection system, and artifacts due to the limited detection view. To address these issues, in this paper we develop a new direct and non-iterative image reconstruction framework using deep learning. Within this approach we combine the dynamic aperture length (DAL) correction algorithm with a deep convolutional neural network (CNN). As demonstrated by simulation and experiment, the resulting DALnet is capable of producing high-resolution projection images of 3D structures in fractions of seconds.

Motivation & Objective

• Address the limitations of standard image formation in photoacoustic tomography using sparse detector arrays.
• Overcome under-sampling, blurring from finite detector response, and view-angle artifacts in projection imaging.
• Develop a direct, non-iterative reconstruction method that enables real-time, high-resolution imaging.
• Integrate dynamic aperture length correction with deep learning to improve image quality and resolution.

Proposed method

• Combine the dynamic aperture length (DAL) correction algorithm with a deep convolutional neural network (CNN) for image reconstruction.
• Use the DAL correction to model the effective aperture size dynamically based on source and detector geometry.
• Train the CNN end-to-end on simulated and experimental data to learn the mapping from raw projection data to high-resolution images.
• Design the CNN architecture to preserve fine spatial details and suppress artifacts from sparse sampling and limited view.
• Apply the trained model directly to new data for real-time, non-iterative image reconstruction.
• Leverage both simulation and experimental validation to ensure robustness and generalization.

Experimental results

Research questions

• RQ1 Can a deep learning framework effectively correct for under-sampling and blurring in photoacoustic projection imaging with sparse detector arrays?
• RQ2 How does combining DAL correction with a CNN improve image resolution and reduce artifacts compared to conventional methods?
• RQ3 To what extent can the proposed method achieve real-time, non-iterative reconstruction while maintaining high image fidelity?
• RQ4 How does the performance of DALnet generalize across different 3D structures and imaging conditions?

Key findings

• DALnet successfully produces high-resolution photoacoustic projection images in fractions of a second, enabling real-time imaging.
• The integration of DAL correction with a CNN significantly reduces artifacts caused by limited detection view and sparse sampling.
• Image quality improvements are demonstrated both in simulation and experimental settings, with enhanced spatial resolution and reduced blurring.
• The method achieves non-iterative reconstruction, avoiding the computational burden of iterative algorithms.
• The framework generalizes well across different 3D structures, as validated through both synthetic and real data.
• The results confirm that the deep learning approach effectively learns complex image reconstruction mappings from raw projection data.