PaGKD

PaGKD

Pairing-free Group-level Knowledge Distillation for Robust Gastrointestinal Lesion Classification in White-Light Endoscopy

Qiang Hu^{1, *}, Qimei Wang^{1, *}, Yingjie Guo¹, Qiang Li¹, Zhiwei Wang^{1, †}

¹ WNLO, HUST
(*: equal contribution, †: corresponding author)

1. Overview

Our framework PaGKD consists of a trainable WLI classifier and a frozen, pretrained NBI classifier. Both classifiers share the same architecture, following ADD (Hu et al. 2025). To facilitate knowledge transfer between unpaired WLI and NBI images, we introduce a group-level feature distillation strategy, instead of regular image-level approaches.

We then apply two complementary modules for multigranularity distillation: Group-level Prototype Knowledge Distillation (GKD-Pro) performs alignment of global classlevel distributions, while Group-level Dense Knowledge Distillation (GKD-Den) aligns local feature.

2. Checkpoints

Model	PICCOLO (AUC)	IH-GC (AUC)	Weights (5-folds)
Ours	0.901	0.840	ckpts
w/o GKD-Pro & GKD-Den	0.712	0.669	ckpts
w/o GKD-Den	0.835	0.755	ckpts
w/o GKD-Pro	0.850	0.773	ckpts

3. Visulization of Results

3.1 ROC Curve:

3.2 t-SNE:

Comparisons between our proposed group-levelm distillation components and their image-level variants.

4. Getting Started

4.1 Recommended Environment:

• Python 3.8+
• PyTorch 2.1+
• TorchVision corresponding to the PyTorch version
• NVIDIA GPU + CUDA
• Install other dependent packages:

cd ADD
pip install -r requirements.txt

4.2 Data Preparation

• Downloading the CPC-Paired dataset (public WLI-NBI paired polyp classification dataset). The file paths should be arranged as follows:

ADD
├── dataset
├── ├── White_light
├── ├── ├── adenomas
├── ├── ├── ├── ├── 01-1.png
├── ├── ├── ├── ├── 02-1.png
├── ├── ├── ├── ├── ......
├── ├── ├── hyperplastic_lesions
├── ├── ├── ├── ├── 011-1.png
├── ├── ├── ├── ├── 011-2.png
├── ├── ├── ├── ├── ......
├── ├── NBI
├── ├── ├── adenomas
├── ├── ├── ├── ├── 01-1.png
├── ├── ├── ├── ├── 02-1.png
├── ├── ├── ├── ├── ......
├── ├── ├── hyperplastic_lesions
├── ├── ├── ├── ├── 011-1.png
├── ├── ├── ├── ├── 011-2.png
├── ├── ├── ├── ├── ......

• Note that the details of dataset splitation in the 5-fold experiment can be downloaded in here. You need to download these .txt files and put them into a newly created folder split and the file paths should be arranged as follows:

ADD
├── split
├── ├── xxx.txt
├── ├── ......

4.3 Training:

Stage 1: pre-traning the NBI classifier:

python train_teacher.py

Stage 2: training the WLI classifier:

python train.py

4.4 Testing and Evaluation:

python test.py

You can also directly download the well-trained model from Google Drive, and predict the results by test.py.

Citation

If you find our paper and code useful in your research, please consider giving us a star ⭐ and citing PaGKD by the following BibTeX entry.

@article{hu2026pairing,
  title={Pairing-free Group-level Knowledge Distillation for Robust Gastrointestinal Lesion Classification in White-Light Endoscopy},
  author={Hu, Qiang and Wang, Qimei and Guo, Yingjie and Li, Qiang and Wang, Zhiwei},
  journal={arXiv preprint arXiv:2601.09209},
  year={2026}
}