MISO: A Controlled Ablation of Masking, Initialization, Sampling, and Optimization for Segmentation in Volumetric Electron Microscopy

Multi-organelle segmentation in volumetric electron microscopy (vEM) faces several challenges, including severe class imbalance, the presence of small, rare classes, and inconsistent class coverage across crops. While recent work has focused primarily on architectural design, the impact of sampling, loss functions, and masking strategies on training effectiveness remains comparatively underexplored in vEM organelle segmentation. Here, we systematically evaluate sampling strategies, loss configurations, masking approaches, and model families (CNNs and vision transformers) on the CellMap benchmark. Using 289 annotated 3D FIB-SEM crops, we establish a 32-class segmentation benchmark with stratified train, validation, and test splits, and evaluate all the methods under the same training and inference settings. Across controlled ablations, the proposed combination of repeat-factor sampling, Tversky-BCE loss, and masking achieved the strongest rare-class performance, increasing rare-class mean Dice (mDice) from 0.3244 under uniform sampling to 0.3409. This corresponds to an absolute gain of +0.0165 mDice and a 5.1% relative improvement, while preserving comparable performance on common classes. Overall, we find that sampling, loss design, and masking contribute as much to performance variation as the choice of architecture, highlighting the importance of training-recipe design alongside model architecture in vEM organelle segmentation.