Physics-consistent denoising of full-field photoacoustic displacement maps using an adaptive residual attention network

Abstract

Full-field photoacoustic tomography (FF-PAT) enables non-contact imaging of optical absorption by mapping the surface displacements induced by photoacoustic waves. In practice, these measured displacement fields often suffer from significant noise, which degrades subsequent image reconstruction. To address this challenge, we propose an Adaptive Residual Attention Network (ARAN) for robust and physics-consistent denoising of FF-PAT displacement maps. Here, the term physics-consistent denotes preservation of the wavefront morphology of photoacoustic waves generated by optical excitation, particularly the sharp gradient transitions associated with acoustic propagation. The network integrates two complementary modules: an Adaptive Residual Fusion Module (ARFM) that dynamically learns to combine residual features according to input noise statistics, and a Residual Self-Recalibration (RSR) mechanism that emphasizes photoacoustic wavefront-consistent structures while suppressing stochastic fluctuations. Unlike conventional CNNs with fixed residual pathways, ARAN adaptively adjusts feature connectivity and attention, achieving superior structural preservation and contrast enhancement. Experimental results demonstrate that ARAN can reconstruct the denoised displacement map from a single noisy input map, which is comparable in quality to those obtained by coherent averaging of 50 input maps. Moreover, ARAN outperforms existing networks such as U-Net and Noise2Noise in both visual and quantitative evaluations, achieving faster inference (0.53 s per frame) suitable for quasi-real time imaging. These results demonstrate a data-driven yet physics-consistent approach to efficient displacement-map denoising, enabling faster and clearer FF-PAT 3D image reconstruction without repeated acquisitions.