Quantitative photoacoustic tomography based on a physics-constrained deep learning framework with implicit priors

Quantitative photoacoustic tomography (qPAT) faces challenges in reconstructing optical absorption coefficients due to the ill-posed nature of the optical inverse problem. This paper presents PIDP-qPAT, a framework that integrates physical models with implicit deep learning priors. The approach incorporates a pre-trained denoising network as an implicit regularizer within an optimization framework, decoupling the reconstruction process into physics-driven data fidelity and learning-driven denoising sub-tasks, which are solved using an alternating optimization strategy. This design eliminates the need for specialized training data by leveraging natural image priors. Comprehensive evaluations through simulations, phantom experiments, and in vivo studies demonstrate that PIDP-qPAT outperforms state-of-the-art techniques in both visual quality and quantitative metrics, while maintaining computational efficiency comparable to purely data-driven methods. This fusion of physical models with data-driven priors provides a robust and practical solution for qPAT and other ill-posed imaging problems.