CRC-ESD gated ABMIL checkpoints

This repository provides the five patient-level cross-validation checkpoints for research on metastasis-risk prediction after endoscopic submucosal dissection (ESD) in colorectal cancer.

Architecture and inputs

Each checkpoint is a PyTorch state_dict for crc_esd.GatedABMILClassifier: a gated ABMIL slide encoder followed by a 1,536 → 256 → 1 classifier. Inputs are UNI-2h H&E patch embeddings at 20× magnification, generated from 256-pixel non-overlapping patches. An input HDF5 file requires a features array with shape (n_patches, 1536); coords is additionally required for Grad-CAM visualisation.

best_model_fold3.pth is the selected checkpoint used in the reported Grad-CAM analyses. The other checkpoints are retained for reproducibility of the patient-level five-fold experiment.

Usage

Install the compatible TRIDENT package and the code package from zzz-pathology/CRC-ESD, then run:

pip install -e '.[visualization]'
crc-esd-predict --features /path/to/slide_features.h5 --checkpoint best_model_fold3.pth
crc-esd-gradcam --features /path/to/slide_features.h5 --checkpoint best_model_fold3.pth --output attention.npz

Intended use and limitations

These parameters are for non-commercial research and reproducibility only. Performance was estimated by internal patient-level five-fold cross-validation in 113 patients (21 metastasis, 92 non-metastasis), with mean AUC 0.918 ± 0.069. The model has not undergone external validation or prospective clinical evaluation and is not a medical device; it must not guide patient management.

No whole-slide images, clinical metadata, or patient-level cellular features are released with these parameters.