A physics-based iterative learning framework for quantitative parametric imaging with application to photoacoustic imaging

Quantitative parametric imaging represents a typical application of knowledge discovery from visual scenes. In numerous instances, purely data-driven neural networks may face challenges in accurately interpreting the physical characteristics of complex visual scenarios, especially when faced with limited or insufficiently diverse training data to encompass the full spectrum of physical phenomena . However, the integration of physical priors into network architectures continuously enhances their capabilities in this domain. This paper introduces a physics-informed iterative learning framework aimed at recovering multiple characteristic parameters of imaging objects from visual scenes. The framework incorporates forward imaging models into the learning process, facilitating mapping from raw measurement signals acquired by the detector to multiple-parameter maps. The application of this framework focuses on the joint quantification of the optical absorption coefficient and the absolute Grüneisen parameter at the pixel level in biological photoacoustic imaging. The iterative gradient network is trained and evaluated on simulation and phantom datasets with established reference values. The results demonstrate significant improvements in the recovery of parametric maps, both qualitatively through visual inspection and quantitatively using evaluation metrics. In a phantom study, the structural similarity and peak signal-to-noise ratio of the recovered parametric maps exhibit improvements of at least 19.5% and 33.5%, respectively, compared to state-of-the-art nonlearning two-step methods. To the best of our knowledge, this represents the first deep learning solution for quantifying absolute Grüneisen parameter from photoacoustic measurements. This work contributes a versatile iterative learning framework for extracting physical insights from detector data.