he-dapi

🚧 Early / work-in-progress (alpha). Only the he2dapi direction is available today; dapi2he (unpaired CycleGAN) is pending training. APIs, weights and outputs may change between releases. Research demo only — not for clinical use.

H&E ↔ DAPI virtual staining — inference library + CLI.

Predicts a virtual DAPI fluorescence image from a brightfield H&E patch (paired Pix2Pix). Weights are downloaded automatically from the Hugging Face model repo.

Install

pip install he-dapi
# optional extras
pip install "he-dapi[wsi]"     # .svs/.tiff support
pip install "he-dapi[onnx]"    # ONNX Runtime

CLI

he-dapi info
he-dapi translate -i he.png -o dapi.png -d he2dapi

Python

from he_dapi import translate, load_generator

pred = translate("he.png", direction="he2dapi")   # -> uint8 HxW numpy array
# or reuse a loaded model across calls:
model = load_generator("he2dapi")
pred  = translate("he.png", model=model)

Large images are processed in 512px tiles (stride 256) with overlap averaging. Input is normalized to [-1, 1] (uint8 / 127.5 - 1); output is [0, 1].

Status

• he2dapi (paired Pix2Pix): available.
• dapi2he (unpaired CycleGAN): added when unpaired training completes.

License

MIT. Research demonstration only — not validated for clinical use.