pkeller01/INSIGHT-preprocessing

INSIGHT preprocessing weights (GrandQC + CRC tissue classifier)

Preprocessing weights used by the HIMICO SDK before feature extraction (QC masks and tissue types). They download automatically when you run the pipeline.

File	Model	Source and licence
grandqc/td_mpp10.safetensors	GrandQC tissue detection (U-Net++, EfficientNet-B0, 10 µm/px)	GrandQC, Weng et al., Nat Commun 2024. CC BY-NC-SA 4.0
grandqc/qc_mpp{1,1.5,2}.safetensors	GrandQC artefact segmentation (U-Net, EfficientNet-B0, 8 classes)	as above
crc_tissue/vit_b16.safetensors	INSIGHT CRC tissue classifier (ViT-B/16, 224 px tiles at 0.5 µm/px, 8 classes: TUM NORM LYM MUS STR DEB MUC ADI)	INSIGHT (Keller et al. 2025). CC BY-NC-ND 4.0

The GrandQC files are the official GrandQC weights (Zenodo 14507273, Zenodo 14041538) converted from pickled PyTorch models to safetensors state dicts, without retraining or other changes. Load them with segmentation_models_pytorch (UnetPlusPlus / Unet, encoder_name="timm-efficientnet-b0", encoder_weights=None).

Non-commercial research use only. The INSIGHT survival model itself is not distributed; it runs on the HIMICO server at Warwick with an API key (see the HIMICO SDK). If you use these weights, please cite GrandQC and INSIGHT:

@article{weng2024grandqc,
  title   = {GrandQC: a comprehensive solution to quality control problem in digital pathology},
  author  = {Weng, Zhilong and others},
  journal = {Nature Communications}, year = {2024}
}
@article{keller2025insight,
  title   = {INSIGHT: Spatially resolved survival modelling from routine histology crosslinked with molecular
             profiling reveals prognostic epithelial-immune axes in stage II/III colorectal cancer},
  author  = {Keller, Piotr and others and Minhas, Fayyaz},
  journal = {arXiv preprint arXiv:2512.22262}, year = {2025}
}