---
license: mit
tags:
  - computational-pathology
  - colorectal-cancer
  - multiple-instance-learning
  - whole-slide-image
  - medical-imaging
language: en
datasets:
  - custom
pipeline_tag: image-classification
---

# CRC-ESD: Metastasis Risk Prediction Model

Trained model weights for **"Deep Learning-Based Analysis of H&E-Stained Histopathological Images for Predicting Metastasis Risk After Endoscopic Submucosal Dissection in Colorectal Cancer"**.

## Model Description

- **Architecture**: FeatherSlideEncoder (gated attention ABMIL) + 2-layer MLP classifier
- **Input**: UNI-2h patch features (1536-d → 512-d embedding) from H&E WSIs
- **Output**: Binary metastasis risk probability
- **Training**: 5-fold stratified cross-validation on 113 CRC patients (294 WSIs)
- **Performance**: Mean AUC = 0.759 ± 0.041

## Files

| File | Description |
|------|-------------|
| `best_model_fold0.pth` | Fold 0 best checkpoint |
| `best_model_fold1.pth` | Fold 1 best checkpoint |
| `best_model_fold2.pth` | Fold 2 best checkpoint |
| `best_model_fold3.pth` | Fold 3 best checkpoint |
| `best_model_fold4.pth` | Fold 4 best checkpoint |

## Usage

```python
import torch
from trident.slide_encoder_models import FeatherSlideEncoder

# Load model
model = FeatherSlideEncoder()
classifier = torch.nn.Sequential(
    torch.nn.Linear(512, 256),
    torch.nn.ReLU(),
    torch.nn.Dropout(0.5),
    torch.nn.Linear(256, 1)
)

checkpoint = torch.load("best_model_fold0.pth", map_location="cpu")
model.load_state_dict(checkpoint["encoder_state_dict"])
classifier.load_state_dict(checkpoint["classifier_state_dict"])
```

## Citation

```bibtex
@article{zhao2025crc_esd,
  title={Deep Learning-Based Analysis of H&E-Stained Histopathological Images for Predicting Metastasis Risk After Endoscopic Submucosal Dissection in Colorectal Cancer},
  author={Zhao, Zhixun and Gong, Changhao and Shi, Yihang and others},
  journal={npj Digital Medicine},
  year={2025}
}
```

## Code

GitHub: [https://github.com/ChanghaoGong/CRC-ESD](https://github.com/ChanghaoGong/CRC-ESD)