Effects of extensive combination of datasets and fine-tuning during training. (A–D) Examples to highlight the effect on the predictive performance of (A, C, and D) mitochondria and (A and B) ER and models trained with data from cells prepared by CF or HPFS, with substantial differences in general appearance and contrast. The images show several comparisons between ground truth annotations and predictions from models trained as described in the insets with data obtained from cells prepared by different sample preparation protocols. Details of the cell and training protocols are in Tables S1, S2, and S8. Voxels corresponding to false positive (cyan arrows) and false negative (red arrows) predictions are indicated. Scale bar, 500 nm. (A) Predictions from cross-domain models, for which the training data and predictions were done using cells prepared with different sample preparation protocols, were less accurate than those obtained from the specialized models, for which training and predictions were done using cells prepared with the same sample preparation protocol. Predictions from the generalist models, obtained by training using ground truth annotations from cells prepared by CF and HPFS, performed only slightly worse than the predictions from the specialized models. (B–D) Effect on the predictive performance of the models by fine-tuning during training.