Aberration Modeling in Deep Learning for Volumetric Reconstruction of Light-Field Microscopy

Optical aberration is a crucial issue in optical microscopes, fundamentally constraining the achievable imaging performance. As a commonly encountered one, spherical aberration is introduced by the refractive index mismatches between samples and their surrounding environments, leading to problems like low contrast, blurring, and distortion in imaging. Light-field microscopy (LFM) has recently emerged as a powerful tool for rapid volumetric imaging. The presence of spherical aberration in LFM will cause substantial changes of the point spread function (PSF) and thus greatly affects the imaging performance. Here, the aberration-modeling view-channel-depth (AM-VCD) network is proposed for LFM reconstruction, effectively mitigating the influence of severe spherical aberration. By quantitatively estimating the spherical aberration in advance and incorporating it into the network training, the AM-VCD achieves aberration-corrected high-speed visualization of 3D processes with uniform spatial resolution and real-time reconstruction speed. Without necessitating hardware modifications, this method provides a convenient way for directly observing the 3D dynamics of samples in solution. The capability of AM-VCD under a large refractive index mismatch is demonstrated through volumetric imaging of a large-scale fishbone of a largemouth bass. Furthermore, the capability of AM-VCD in nearly 100 Hz volumetric imaging of neutrophil migration and a beating heart in living zebrafish is investigated.

Automated summary

Optical aberration is a common issue in optical microscopes and can lead to low contrast, blurring, and distortion in imaging. Light-field microscopy (LFM) has recently been used for rapid volumetric imaging, but spherical aberration affects its imaging performance. To address this, the aberration-modeling view-channel-depth (AM-VCD) network is proposed for LFM reconstruction, achieving aberration-corrected high-speed visualization of 3D processes. This method enables convenient observation of the 3D dynamics of samples in solution without hardware modifications.