Three-Dimensional Reconstruction of Serial Block-Face Scanning Electron Microscopy Using Semantic Segmentation based on Semi-Supervised Deep Learning.

Serial block-face scanning electron microscopy (SBF-SEM) is employed to achieve high-resolution volume reconstructions and detailed ultrastructural analyses of complex organelles. The performance of SBF-SEM is evaluated according to the accuracy of segmentation. Our study introduces a semi-supervised learning approach using a segment interpolation method to mitigate the costs of manual segmentation. The shapes and locations of individual segments between sparsely annotated label images are estimated using the proposed method. The proposed method is particularly well suited for SBF-SEM, where alignment and fine cutting of samples allow for accurate predictions with a minimal amount of labelled data. To validate the deep neural networks trained using the proposed method, the F-1 score metric and the K-fold technique were utilized. The results achieved an F-1 score of 0.89 for mouse brain cells and 0.84 for inverted images during the validation process for semi-supervised learning. Testing on an independently separated test dataset yielded scores of 0.84 for mouse brain cells and 0.80 for inverted cases. The automatically segmented results were then reconstructed in volume images using the marching cube algorithm. This allows for a three-dimensional (3-D) analysis of complex organelles, with potential applications in the fields of biology and medicine.