RC-24 Rare Tumor Histopathology Patch Dataset

This dataset supports the study “VitaminP: cross-modal learning enables whole-cell segmentation from routine histology.”

RC-24 is a curated rare cancer validation dataset comprising 24 paired hematoxylin and eosin (H&E) and multiplex immunofluorescence (mIF) histopathology image patches (512 × 512 pixels) acquired at 20× magnification. Each pair represents a distinct tumor subtype not observed during training of the VitaminP segmentation framework.

The cohort spans diverse anatomical systems, including gastrointestinal/hepatobiliary, genitourinary/gynecologic, thoracic/mesothelial, and endocrine/soft tissue malignancies, enabling evaluation of rare morphology generalization and cross-modal consistency under out-of-distribution conditions.

All H&E patches were annotated by expert pathologists. Nuclei were exhaustively delineated across each tile, while whole-cell annotations were conservatively provided for high-confidence cells with clearly identifiable cytoplasmic boundaries. This partial whole-cell annotation strategy was adopted to avoid ambiguous cytoplasmic boundary labeling in morphologically complex regions.

The dataset contains 11,798 manually delineated instances:

7,646 nuclei

4,152 high-confidence whole-cell instances

Each H&E patch is paired with a spatially aligned mIF image, enabling cross-modal segmentation validation and molecular-boundary consistency analysis.

Case-level metadata, including diagnosis and anatomical category, are provided in RC-24_metadata.csv. Detailed annotation protocol and tumor subtype information are documented in the accompanying README file.

RC-24 was used for rare morphology and cross-modal validation in the VitaminP study.