[MICCAI 26] MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models

MedPruner is a training-free hierarchical visual token pruning method for efficient 3D medical image understanding in vision-language models. It removes redundant inter-slice information with anchor-based filtering, then performs token-level dynamic information nucleus selection to reduce inference cost while preserving diagnostic accuracy.

This repository also includes Med3D, an evaluation framework for 3D medical imaging (CT/MRI) vision-language models, with built-in MedPruner support.

✨ Key Features

• 3D Medical Image Evaluation: End-to-end evaluation on 3D volumetric data including CT and MRI (Amos, M3D, 3DRad)
• Visual Token Pruning (MedPruner): Two-level pruning strategy combining slice-level and token-level compression to reduce visual tokens, memory usage, and inference time
• Multi-Model Support: Unified interface for Qwen3-VL, HuluMed, and MedGemma medical VLMs
• Multi-GPU Parallelism: Data-parallel evaluation across multiple GPUs

🛠️ Setup

# Create conda environment
conda create -n med3d python=3.10
conda activate med3d

# Video processing dependencies
pip install decord ffmpeg-python imageio opencv-python
pip install qwen_vl_utils nltk rouge mathruler pylatexenc

# 3D medical image processing (NIfTI files)
pip install nibabel

# Install other dependencies
pip install -r requirements.txt

# Install medpruner package
cd medpruner && pip install -e .

📚 Datasets

Dataset download: Hulu Eval Dataset

Place the datasets under data/Eval/ with the following structure:

data/Eval/
├── Amos/
│   ├── amos_test_crop_32/
│   └── amos_val_mrg_sft.json
├── 3DRad/
│   └── ...
└── M3D/
    └── ...

🤖 Supported Models

Model	Backend	MedPruner
Qwen3-VL	transformers	✓
HuluMed Qwen2	transformers	✓
MedGemma	transformers	✓

🧪 Supported Benchmarks

Dataset	Modality	Description
Amos	3D CT/MRI	Multi-organ segmentation and report generation
M3D	3D CT	Full-body CT visual question answering
3DRad	3D CT	Radiology report generation

⚙️ MedPruner Configuration

MedPruner reduces visual tokens through two pruning levels:

• Slice-level pruning (slice_compose): Drops redundant frames whose anchor-frame L1 distance is below threshold gamma
• Token-level pruning (token_compose): Retains key tokens whose cumulative attention mass reaches tau, and aggregates the rest by similarity

Configuration file config/config.json:

{
  "medpruner": {
    "slice_compose": true,
    "token_compose": true,
    "tau": 0.9,
    "gamma": 0.05
  }
}

Parameter	Type	Description
slice_compose	bool	Enable slice-level inter-frame compression
token_compose	bool	Enable token-level attention selection
tau	float	Cumulative attention mass threshold (0~1), controls token retention. Lower values keep fewer tokens
gamma	float	Anchor-frame L1 distance threshold; frames below this value are discarded

Pruning rates: after evaluation, compression_rate (slice pruning retention), dyn_token_rate (token pruning retention), and select_rate (overall retention) are reported.

🚀 Evaluation

cd MedUniEval

# Evaluate Amos / 3DRad
bash ./eval.sh

# Evaluate M3D (chunked evaluation)
bash ./eval_M3D_chunked.sh

Key environment variables:

Variable	Description
EVAL_DATASETS	Dataset name, e.g. Amos, M3D, 3DRad
MODEL_NAME	Model registry name, e.g. Qwen3-VL, Hulumed_qwen2
MODEL_PATH	Path to model weights
COMPRESSION	Pruning method, set to medpruner to enable
CUDA_VISIBLE_DEVICES	Available GPUs; multiple GPUs enable data parallelism

🙏 Acknowledgements

Greatly appreciate the tremendous effort for the following projects!

• Visionzip
• Hulu-Med
• MedGemma

📖 Citation

@misc{liu2026medprunertrainingfreehierarchicaltoken,
      title={MedPruner: Training-Free Hierarchical Token Pruning for Efficient 3D Medical Image Understanding in Vision-Language Models}, 
      author={Shengyuan Liu and Zanting Ye and Yunrui Lin and Chen Hu and Wanting Geng and Xu Han and Bulat Ibragimov and Yefeng Zheng and Yixuan Yuan},
      year={2026},
      eprint={2603.11625},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.11625}, 
}