Pipelines used for training and deep-learning neural network prediction. Schematic representation of the deep-learning approach for recognizing intracellular structures in FIB-SEM volume images using a 3D U-net encoder-decoder neural network. (A) For training, three-dimensional stacks containing FIB-SEM data, augmented as described in Materials and methods, are provided as input images to the 3D U-Net; in this example, the stack includes a limited number of three-dimensional ground truth annotations for the ER in the form of binary masks (yellow). The ER predicted by the 3D U-net model is a 3D probability map, whose error is calculated by comparing the ground truth annotation with the cross-entropy loss. The model parameters are iteratively updated during training until convergence of the cross-entropy loss is achieved. (B) For prediction, small 3D stacks with data not used for training covering the complete FIB-SEM volume image of a naïve cell (or from the remaining regions of the cell used for training) are provided as input to the 3D U-net model trained in A. In this example, the predicted ER is a thresholded 3D probability map for the entire cell volume.