Performance of the deep-learning network to predict in naïve cells. (A and B) Visual comparisons between predictions (crimson) by 3D U-net models trained using combined data from two HEK293A cells to recognize (A) ER or (B) Golgi apparatus and corresponding ground truth annotations (blue) in the naïve BSC-1 and SUM 159 cells not used for training (Table S1). The representative images of single plane views from FIB-SEM volume data are from cells prepared by CF isotropically acquired at a 5 nm resolution; red and yellow arrows highlight small regions containing voxels of false positive and false negative assignments. Scale bar, 500 nm (see Videos 1 and 2).