Light field microscopy (LFM) is widely employed in real-time cellular activity observation, three-dimensional tissue structure imaging, and organ pathological diagnosis. However, the quality of light field microscopic images is often compromised by inherent lens defects and sample-induced optical aberrations due to variable refractive index distributions. Current aberration correction methods primarily exploit the intensity information of the object, ignoring latent sample phase image data such as thickness variations and 3D morphology. Thus, we introduce a phase-intensity dual-path network (PCANet) designed for high-resolution reconstruction in light field microscopic aberration correction and adopt deep learning to decouple two-dimensional light field microscopic intensity and phase information for enhanced resolution. Experimental results indicate that this deep learning approach effectively replaces light field digital adaptive optics, and achieves aberration correction, high-resolution image reconstruction, and restoration of sample detail edges, thereby recovering the resolution and signal-to-noise ratio of light field microscopic imaging.

We propose a PCANet that combines multi-dimensional light field data with a deep learning model to correct imaging aberrations and perform high-resolution reconstruction. The model consists of two serially processed network segments that handle original low-resolution aberrated light field data, ultimately outputting high-resolution reconstruction via light field microscopic decoupling and PCANet modules. This reduces reliance on complex aberration compensation devices, enabling cost-effective and high-resolution light field microscopic reconstruction. The light field microscopic imaging system captures the original low-resolution aberrated data, which is then decoupled by the light field decoupling module into intensity and phase information. The PCANet extracts features from these dimensions, fusing and mining the two-dimensional sample information to enhance aberration correction and achieve high-resolution reconstruction without hardware compensation. Thus, our deep learning model which requires only low-resolution aberrated light field data as input and outputs high-resolution aberration-corrected images significantly simplifies computation and exhibits superior reconstruction quality in experimental results.

The US Air Force standard USAF is adopted to verify the aberration correction capabilities of PCANet. Reconstruction results show that while the original light field aberration image barely resolves the fifth group of element 2 (line width is 13.92 μm) at the edge, the digital adaptive optics (DAO) method aberration correction reaches the sixth group of element 6 (line width is 4.38 μm). Our process restores the seventh group of element 5 (line width is 2.46 μm), indicating effective aberration correction and high-resolution reconstruction, and near-accurate levels regardless of significant distortion in light field microscopic edges or lesser aberration influences in central areas. Introducing phase information enhances network aberration correction, which outperforms image super-resolution network (VDSR) and Richardson-Lucy deconvolution algorithm (R-Lucy) in horizontal comparisons. Meanwhile, higher peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) performance metrics corroborate the efficacy of our proposed network.

We present an innovative application of deep learning technology to light field microscopic aberration correction, with microscopic samples’ intensity and phase information employed. By conducting resolution plate experiments and tests on egg embryo slices and onion epidermal layers, we demonstrate that the original light field aberration data can be effectively corrected via network recovery to surpass DAO aberration correction methods and R-Lucy deconvolution in terms of reconstructed image resolution and clarity. By decoupling and integrating phase and intensity feature information, our approach avoids complex iterative calculations and additional physical devices, simplifies operations, and reduces system complexity and cost, with potential for practical application advancement.