PIT-Net: Physics-Informed Transformer with Differentiable Diffusion Constraints for Quantitative Photoacoustic Tomography of Deep Tissues

Quantitative photoacoustic tomography (QPAT) faces the ill-posed inverse problem of decoupling optical absorption from light fluence. Standard deep learning models often lack physical consistency and struggle to generalize in vivo. To address this, we propose PIT-Net, a physics-informed transformer framework for multi-wavelength QPAT. PIT-Net integrates a Swin Transformer backbone to extract multiscale, cross-spectral features with an embedded differentiable physics propagator (DPP). During the forward pass, the DPP solves the steady-state photon diffusion equation under Robin boundary conditions, coupled with an acoustic residual feedback loop to promote physical consistency. Trained entirely on synthetic data, PIT-Net demonstrates direct transferability to in vivo targets without fine-tuning. It achieves high quantitative accuracy in simulations (PSNR: 45.475 ± 3.599 dB; RMSE: (6.0 ± 2.7) × 10⁻³) and reconstructs high-resolution absorption images across physical phantoms and in vivo tissues. Longitudinal monitoring of murine tumors further validates PIT-Net as a robust, physically interpretable solution for QPAT.