End-to-end aberration correction network for enhancing miniature microscope resolution

Miniature microscopes are widely used for in vivo neural activity observation and point-of-care medical diagnosis, but often suffer from low resolution due to uncorrected optical aberrations. Designs that fully correct aberrations typically involves complex and expensive custom aspheric lenses. While image post-processing using measured point spread functions can effectively correct aberrations, these measurements are time-consuming, laborious and have high uncertainty from alignment errors. Given these issues, we propose an accurate Point Spread Function (PSF) modeling with geometric ray tracing and wavelet coherent superposition, accounting for most microscopic optical aberrations. Based on this, we develop an efficient image synthesis method, leveraging the rotational symmetry of optical systems to reduce PSF computation for a space-variant, multi-patch convolution imaging model. Additionally, we create bright-field and dark-field microscopy datasets for training the aberration correction network. The proposed deep learning method is based on multiple Wiener deconvolution and multi scale network, effectively correcting shift-variant, wavelength-dependent aberrations and significantly improving the image resolution. We validated our methods on two miniature microscope prototypes. Although these systems only used two conventional off-the-shelf lenses, our proposed joint aberration correction network improved the resolution to recognize 0.7 mu m chart bar with a long working distance of 7 mm. This approach provides a lowcost, high-efficiency solution for enhancing the performance of miniature microscope systems.