iSight-cell: cell-level assessment of immunohistochemistry staining

The cell-level model of iSight. Every cell in an IHC image is segmented, iSight-target selects the cells of interest for the tissue, and iSight-cell scores each selected cell for staining intensity (negative / weak / moderate / strong) and subcellular location (none / nuclear / cytoplasmic-membranous / both). Image-level results are built up from the cells, giving a cell count, a spatial map and a measured stained fraction.

iSight-cell requires zhihuanglab/iSight-target.

Model

UNI2-h backbone (ViT-g/14), fully fine-tuned, with two classification heads. Input is a 64×64 RGB crop centred on the cell, resized to 224 and ImageNet-normalised.

Files

config.json                          model configuration (fetch this with the checkpoint)
checkpoints/iSight-cell.pt   weights (model state dict only)

from huggingface_hub import hf_hub_download
cfg  = hf_hub_download("zhihuanglab/iSight-cell", "config.json")
ckpt = hf_hub_download("zhihuanglab/iSight-cell", "checkpoints/iSight-cell.pt")

Links

• Code: github.com/zhihuanglab/iSight (isight_cell/)
• Image-level model: zhihuanglab/iSight-slide
• Training dataset: nirschl-lab/hpa10m

License

PENN Academic Software License Agreement: non-commercial research use only.

Contact

Zhi Huang — zhi.huang@pennmedicine.upenn.edu