VitaminP: Processed Zarr Datasets for Cross-Modal Whole-Cell Segmentation in Histopathology

This dataset release contains processed Zarr-format inputs used for training and evaluating VitaminP, a cross-modal learning framework for whole-cell segmentation in histopathology. The package includes processed H&E and/or multiplex immunofluorescence-derived image tiles and corresponding instance masks organized in the same structure expected by the public VitaminP dataloader.

The release contains processed Zarr archives for ORION-CRC, Xenium, TissueNet, PanNuke, Lizard, MoNuSeg, TNBC, NuInsSeg, CryoNuSeg, BCData, CoNSeP, MoNuSAC, Kumar, and CPM17. A metadata archive is included with the YAML configuration, sample-level manifest, patch-level final train/validation/test split manifest, and exact Xenium train/validation split indices.

Final patch counts:
Train: 10,986
Validation: 1,740
Test: 2,415

The validation split keeps existing public-dataset validation partitions and adds a 10% patch-level validation split from the Xenium training pool. The test split includes held-out benchmark test sets and lungV1 as a held-out Xenium test sample.

The public code repository should be used together with this dataset release to reproduce the dataloader behavior and experiments.