SlideCheck Pre-trained Models

Official pre-trained models for SlideCheck - a lightweight dual-head patch classifier for computational pathology.

Available Models

Model	Description	BRACS Cancer AUC	Mean BACC	Size	Download
phase1_baseline.pt	Base model (Mixup+FP)	0.9345	0.8333	9.8 MB	📥
v4a_bl_ft.pt	Co-evolution FT (best overall)	0.9253	0.8591	9.8 MB	📥
v5a_bl_scr.pt	Co-evolution Scratch (best scratch)	0.9170	0.8555	9.8 MB	📥
v6c_bl_ft.pt	V3c confidence weighting	0.9322	0.8508	9.8 MB	📥

Quick Start

Installation

pip install huggingface_hub torch numpy h5py
git clone https://github.com/lingxitong/SlideCheck.git
cd SlideCheck
pip install -e .

Download and Use

from huggingface_hub import hf_hub_download
from slidecheck.inference import SlideCheckPredictor

# Download model (choose one)
model_path = hf_hub_download(
    repo_id="infamousgxy/slidecheck",
    filename="phase1_baseline.pt"  # or v4a_bl_ft.pt, v5a_bl_scr.pt, v6c_bl_ft.pt
)

# Load predictor
predictor = SlideCheckPredictor(model_path)

# Predict from H5 file (TRIDENT format)
results = predictor.predict_from_h5('features.h5')

print(f"Abnormal probability: {results['prob_abn'].mean():.4f}")
print(f"Cancer probability: {results['prob_can'].mean():.4f}")

Model Comparison

Cross-Domain Performance

Model	BRACS Can AUC	UNITO Can AUC	CAMEL Can AUC	Mean BACC
Phase1 BL	0.9345	0.7977	0.8517	0.8333
V4a BL FT	0.9253	0.8281	0.9073	0.8591 ⭐
V5a BL SCR	0.9170	0.8257	0.9025	0.8555
V6c BL FT	0.9322	0.8172	0.8843	0.8508

Recommendation:

• Best overall: V4a BL FT (highest mean BACC across 3 datasets)
• Best single-dataset: Phase1 BL (highest BRACS Cancer AUC)
• Best scratch: V5a BL SCR (no Phase1 dependency)

Model Details

Architecture

All models share the same architecture:

SlideCheckMLP:
  LayerNorm(2560)
    → Linear(2560, 768) → GELU → Dropout(0.1)
    → Linear(768, 768) → GELU → Dropout(0.1)
    → head_abn: Linear(768, 1)  # Normal/Abnormal
    → head_can: Linear(768, 1)  # Cancer/Non-Cancer

• Input: Virchow2 features (2560d = CLS 1280d + MeanPool 1280d)
• Parameters: 2,564,098
• Output: Two logits (abnormal, cancer)

Training Data

Dataset A (859,379 patches from 8 datasets):

• CRC-100K, ESCA_TCGA, RenalCell
• Spider_Breast, Spider_Colorectal, Spider_Skin, Spider_Thorax
• TCGA_Uniform

Co-evolution models (V4a, V5a, V6c) additionally use:

• Pseudo-labels from MIL mining (WSI-level supervision)
• Negative bag enhancement (V4a only)

Input Format

SlideCheck requires Virchow2 features (2560d). See phase1_baseline_README.md for feature extraction code.

H5 Format (TRIDENT)

# One h5 file per WSI:
{
    'features': [N_patches, 2560],   # float32
    'coords':   [N_patches, 2],      # int64
}

Citation

If you use these models in your research, please cite:

@article{zhu2025subspecialty,
  title={Subspecialty-specific foundation model for intelligent gastrointestinal pathology},
  author={Zhu, Lianghui and Ling, Xitong and Ouyang, Minxi and Liu, Xiaoping and Guan, Tian and Fu, Mingxi and Cheng, Zhiqiang and Fu, Fanglei and Zeng, Maomao and Liu, Liming and others},
  journal={arXiv preprint arXiv:2505.21928},
  year={2025}
}

License

Apache License 2.0

Links

• GitHub Repository: https://github.com/lingxitong/SlideCheck
• Original HuggingFace: https://huggingface.co/xtxx/SlideCheck

Model Cards

• phase1_baseline_README.md - Base model details
• v4a_bl_ft_README.md - V4a co-evolution finetune
• v5a_bl_scr_README.md - V5a co-evolution scratch
• v6c_bl_ft_README.md - V6c confidence weighting