Foreground-Seeded Patch Sampling for 3D Multi-Organelle Segmentation in Volume EM

Training 3D segmentation models on volume electron microscopy (vEM) requires converting large annotated volumes into local image--label examples. How these examples are constructed determines which structures are sampled, how much surrounding tissue is visible, and which labels co-occur during optimization. We study this problem for multi-organelle segmentation using CellMap eFIB-SEM data. A foreground-seeded sampling pipeline first stores locations drawn from annotated structures, then applies bounded spatial jitter during training and inverse-frequency sampling over ten foreground seed categories. Within this fixed sampling pipeline, we compare (64^3) and (128^3) patch extents, repeat both configurations with geometric and intensity augmentation, and evaluate a signed-distance-transform (SDT) formulation at (64^3). Each patch extent is used consistently during training and inference. All models are scored on identical central (64^3) regions from 73 locations in a crop excluded from training and validation. We find that increasing patch size produces the largest performance change among the configurations studied: categorical mean intersection over union (IoU) rises from 0.279 at (64^3) to 0.471 at (128^3), while augmented models increase from 0.158 to 0.427. Under the tested settings, geometric and intensity augmentation does not improve the corresponding categorical configurations, whereas SDT prediction reaches 0.313 mean IoU at (64^3) and changes class-specific behavior. These results suggest that patch extent should be considered explicitly in foreground-seeded volumetric segmentation pipelines.