Sparse scanning encoding and neural network decoding for compressed photoacoustic microscopy

Abstract

Photoacoustic microscopy (PAM) offers high-resolution, non-invasive, and label-free imaging, making it invaluable for biomedical research. However, slow data acquisition and high sampling requirements remain key challenges that limit its broader applicability and scalability. We propose an Information-Efficient Photoacoustic Microscopy (IE-PAM) that jointly integrates sparse scanning encoding with neural network decoding to achieve high-quality reconstruction from extremely limited measurements. Specifically, IE-PAM employs a sparse-scanning acquisition scheme guided by random binary masks and reconstructs high-fidelity images using AFDU-Net, a custom-designed neural decoder trained on fully sampled ground truth data. Our system can faithfully recover detailed anatomical structures from as little as 1.5 % of the full sampling rate, corresponding to more than a 66-fold increase in acquisition efficiency. In in-vivo experiments on mouse ear vasculature, IE-PAM outperforms both traditional and learning-based baselines in fine vascular fidelity, artifact suppression, and robustness across varying sampling rates. By minimizing information redundancy at the acquisition stage and enabling accurate reconstruction from minimal data, IE-PAM provides a foundation for efficient, fast and scalable photoacoustic imaging in both preclinical and research applications.