MUnet: A model-based unrolled reconstruction framework for computational photoacoustic imaging

Photoacoustic imaging provides unique optical contrast with ultrasonic penetration, enabling high-resolution visualization of biological structures in vivo. However, robust image reconstruction remains challenging under practical conditions such as limited-view detection, sparse sampling, and strong noise. Existing approaches face trade-offs between efficiency, accuracy, and physical interpretability. To address these limitations, we propose MUnet, a model-based unrolled architecture that integrates physics-informed data-fidelity term with supervised image-domain enhancement module. Instead of learning an explicit regularization functional or its proximal operator, this design alternates non-trainable, analytic updates with neural image refinement, combining the interpretability of model-based methods with the flexibility of deep learning. Experiments on in vivo PAI data demonstrate that MUnet consistently outperforms both traditional and learning-based approaches, delivering superior image quality and substantially reduced computation time. MUnet thus represents a promising solution for rapid and high-fidelity PAI, advancing its potential for real-time clinical applications.