Physics-Guided Deep Learning for 3D Photoacoustic Microscopy

Photoacoustic microscopy (PAM) achieves high lateral resolution through tight light focusing but suffers from a narrow depth of focus (DOF). Here the enhancement of PAM image quality for 3D microscopy is aimed by achieving high resolution over a large DOF, accurately restoring object sizes, and minimizing artifacts introduced during image acquisition. A novel approach is proposed that initially acquires 3D PAM images with a large DOF but low resolution through loose light focusing, and subsequently enhances resolution and image quality using a deep learning network termed LDHR-Net. This network is trained on synthetic low-resolution and high-resolution image pairs generated using a physical Gaussian beam model, which accounts for depth-dependent blurring. Application of the proposed model to experimentally acquired phantom data demonstrates significant improvements, achieving lateral resolution of approximate to 4 mu m across an unprecedentedly large DOF of approximate to 4.5 mm, a 25-fold increase compared to the intrinsic DOF (approximate to 0.18 mm) of a PAM system with the same lateral resolution. The model's effectiveness and robustness are further validated qualitatively and quantitatively on in vivo microvasculature structures, including those in the mouse ear, back, and brain. This method provides an effective and practical solution for obtaining high-quality 3D PAM images.