DLAO: a physics-informed deep learning framework for aberration correction in optical coherence tomography

Optical coherence tomography (OCT) is an indispensable high-resolution, non-invasive imaging modality in biomedical diagnostics and research. Its performance, however, is frequently compromised by optical aberrations arising from system imperfections and sample-induced inhomogeneities, which degrade image quality by reducing spatial resolution and obscuring fine details. To address this challenge, we introduce a physics-informed deep learning adaptive optics (DLAO) framework for the efficient correction of complex aberrations in OCT images. The framework introduces a pseudo-point spread function (pseudo-PSF) preprocessing step, which reformulates the high-dimensional image restoration task into a low-dimensional physical parameter estimation problem. This is achieved by processing pairs of images captured with known bias aberrations, which effectively decouples sample-specific structural information from the aberration features at the input stage, thereby reducing data dimensionality and computational load. To decode these compact aberration features, we designed the layerwise adaptive progressive attention network (LAPANet), which integrates a multi-scale feature fusion mechanism with a novel LAPA module to enhance the capture of hierarchical, multi-scale features. This synergistic design enables the network to precisely reconstruct critical image regions, demonstrating exceptional performance in restoring high-frequency details and modeling complex aberration patterns. Experimental validation on tasks involving the correction of aberrations defined by Zernike coefficients demonstrates that our architecture outperforms several mainstream deep learning models by achieving higher peak signal-to-noise ratio and structural similarity index scores, while maintaining high inference efficiency. Furthermore, ablation studies and a series of generalization and extension experiments were conducted to validate the effectiveness, robustness, and generalization capability of the proposed architecture, confirming its efficacy and practical value in enhancing OCT image quality.