Fig. 2

Experimental setup of our workflow. A Orderly serial section acquisition. Slice the tissue block with an ultrasonic oscillating diamond knife and collect sections orderly on a formvar-coated slotted grid. B Locating ROIs in first two sections. Locating the ROI in the first two sections manually, and take 5K (magnification is 5000) and 20K (magnification is 20000) images of ROI as the reference for other sections. C Rough locating of ROI in other sections. Use affine transformation and the position relationship between ROIs to locate the ROI at low magnification (5K). D Precise locating of ROI in other sections. Manually determine the precise location (20K) of ROI based on coarse locating. E Coarse alignment in multi-resolution images. Make full use of the multi-resolution images obtained during the workflow, and use the feature point-based approach to align the ET volumes coarsely. F BP-Elastic fine alignment. Use BP-Elastic to fine align the high-magnification images, then apply the calculated deformation fields to the deformed ET volumes from E. G Missing information estimation. Estimate the number of lost images between the adjacent surfaces of the volumes. H Missing information generation. Use a frame interpolation approach based on deep-learning network to generate the missing information. I Target structure segmentation. Use a 3D neural network combined with manual proofreading to segment the target structure