Deep learning framework for rapid aberration correction in reflection matrix microscopy

Aberrations caused by angle-dependent phase distortions in scattering media degrade image quality in optical imaging. In this study, we present a deep learning-based aberration correction method for reflection matrix microscopy that iteratively corrects both input and output aberrations. By leveraging the structural consistency of output aberrations across different incident wavevectors, we train a U-Net-based model to predict and correct aberrations directly from the reflection matrix. Our iterative inference and correction process effectively eliminates round-trip aberrations, significantly enhancing imaging quality. To further improve efficiency, we introduce a covariance matrix-based training strategy, eliminating the need for explicit input aberration correction and reducing iteration time by half. Our approach achieves a 100-fold computational speedup over conventional wave correlation-based algorithms while maintaining high correction accuracy. We validate our method through numerical simulations and experimental data, demonstrating robustness across various aberration conditions. This deep-learning framework enables real-time, label-free imaging, overcoming computational bottlenecks in aberration correction and paving the way for rapid, high-resolution imaging in biomedical applications where real-time aberration correction is essential.