measty/himico_CRC_he2ihc

CRC H&E to multiplex IHC (HIMICO)

A model built on UNIStainNet that translates H&E-stained colorectal cancer (CRC) tissue into RGB images of multiplex IHC (mIHC; CDX2, MUC2, MUC5 and CD8 with haematoxylin). They were trained on paired H&E and mIHC patches from the HIMICO dataset. The SPADE U-Net generator is conditioned on a 32 × 32 grid of patch features from a pathology foundation model.

Models

Folder	Foundation model	Name in the inference script
uni2-h/	UNI2-h, frozen	crc-himico-uni2 (default)

Each folder contains an exported model:

• model.safetensors: the EMA generator weights.
• config.json: the architecture hyperparameters and inference settings. Tiles are 512 × 512 pixels, read at 0.5 µm/pixel.

The foundation model's weights are not included.

Usage

Whole-slide inference uses scripts/infer/infer_crc_himico_wsi.py from MyUNIStainNet, which downloads this model by default:

python scripts/infer/infer_crc_himico_wsi.py \
    --wsi "/path/to/slides/*.svs" \
    --output "/path/to/output/*_mihc.ome.tiff"

The output is a pyramidal OME-TIFF at the slide's resolution. To pick a model explicitly, pass a name from the table (--model crc-himico-uni2) or a folder of this repository (--model measty/himico_CRC_he2ihc/uni2-h --revision v1.0).

The models also need UNI2-h, which is gated: request access on its model page, then log in with hf auth login or set HF_TOKEN. Downloads are cached, so machines without internet access can run with HF_HUB_OFFLINE=1 once both models are cached.

Training

The uni2-h model was trained with a frozen UNI2-h encoder

License and terms of use

These models are released under CC-BY-NC-ND 4.0, for non-commercial academic research only. They were trained on UNI2-h outputs, so they count as derivatives of UNI2-h under its terms of use, which prohibit commercial use without prior approval. They are not intended for clinical use.

Citation

If you use these models, please cite UNIStainNet (arXiv:2603.12716) and UNI:

Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine (2024). https://doi.org/10.1038/s41591-024-02857-3