LSMS-Net: Resolving multiple scattering via physics-constrained deep learning for photoacoustic imaging in porous media

Photoacoustic imaging of porous heterogeneous media is fundamentally challenged by spatially distributed optical absorbers that simultaneously serve as acoustic scatterers. This dual role causes multiple scattering, wavefront distortion, and frequency-dependent attenuation, which cannot be decoupled by conventional homogeneous-medium reconstruction algorithms. As a result, deep-tissue images suffer from severe image blurring and irreversible loss of high-spatial-frequency microstructural details. To address this, we propose a physics-constrained deep learning framework termed the Layer-Stripping Multi-Scattering Network (LSMS-Net), which recursively solves the inverse multiple scattering problem. LSMS-Net implements a top-down, curriculum-driven strategy to progressively estimate propagation-induced effects and correct raw measurements layer by layer. The framework integrates a Ghost Network for non-local reverberation, an Adaptive Scattering Point Spread Function Layer for wavefront distortion, and a Spectral Consistency Constraint for frequency-dependent attenuation. Quantitative evaluation via numerical simulations and phantom experiments confirms accurate, high-fidelity reconstruction of deep porous structures under strongly scattering, showing potential for clinical application in bone-related diseases.