Improved shape from focus network for extended depth of field and rapid three-dimensional reconstruction biological microscopy

Abstract

Three-dimensional (3D) reconstruction microscopy has played an important role in advancing the elucidation of the roles and structures of biological cells, but the current mainstream optical microimaging techniques make it difficult to capture the 3D structures of dynamic organisms. Therefore, this paper proposes a fast and versatile end-to-end improved shape-from-focus (ISFF) network and enlarged selective kernel (ESK) module, which are applied to obtain microscopes with a large depth of field and high-precision 3D reconstruction capability. To characterize the feasibility of ISFF, our algorithm achieves a higher quality of the fused image compared to six prevalent deep learning image fusion algorithms. We apply our microscopy to 3D imaging of live biological samples such as bee tentacles, C.elegans, and zebrafish without fluorescent labels or anesthesia, and our experimental results show that high-resolution 3D observation of biodynamic processes can be achieved in 1.86 s. Quantitative analysis of the interface between the standard gauge blocks shows that the microscopy’s depth of field extends to 1200 μm under a 10 × objective and the relative errors of the reconstruction for the two gauge blocks are 0.61 % and 0.54 %, respectively. Our network eliminates the need to train exclusively on model organisms such as C. elegans and zebrafish, while still achieving good 3D reconstruction results. It not only expands the application range and system robustness of biomicroscopy but also provides new perspectives and tools for living model organisms at the millimeter scale.