Uncertainty-Aware Semantic Segmentation of Serial Block-Face Scanning Electron Microscopy Images and Three-Dimensional Analysis Method Using Reconstructed Segmentation and Uncertainty Volumes.

Recent advances have leveraged serial block-face scanning electron microscopy (SBF-SEM) images and deep neural network models for automatic semantic segmentation, enabling three-dimensional (3D) analyses of cellular organelles. However, applying such models in real-world scenarios raises reliability concerns, highlighting the need for more trustworthy models. This study introduces uncertainty-aware and quantifiable models based on the deep ensemble (DE) method for the semantic segmentation and 3D reconstruction of SBF-SEM images. We trained these models to produce both accurate segmentations and well-calibrated uncertainty estimates, thereby enhancing their reliability. We analyzed segmentation and calibration performance across various configurations and derived empirical insights into unexpected behaviors in the DE method, yielding practical implications for building uncertainty-aware and quantifiable models. Moreover, we reconstructed not only the 3D semantic segmentation volume but also the corresponding 3D uncertainty volumes. By leveraging these volumes, we propose a novel 3D analysis method that addresses reliability concerns and can support informed decision-making in SBF-SEM applications.