LaGuadia: Language-Guided Adaptive Distillation from Pathology Foundation Models [MICCAI 2026]

Gangsu Kim, and Won-Ki Jeong†, HVCL@KU
† Corresponding Author

Overview

LaGuadia (Language-Guided Adaptive DistillAtion), a framework that develops a compact pathology image encoder by dynamically integrating expertise from multiple PFMs under clinical linguistic guidance

⚙️ Installation

0. Inatall CLAM

We use CLAM, integrated within TRIDENT, for tissue segmentation and patching.

git clone https://github.com/mahmoodlab/trident.git && cd trident
pip install -e .

1. Install dependencies

git clone https://github.com/hvcl/LaGuadia.git && cd LaGuadia
pip install -r requirements.txt

🔥 Training

Note

For efficient training, pre-extracting teacher features before training is highly recommended.

Stage 1. Keyword Extraction

Before training, keyword extraction from pathology reports must be performed.
Keyword extraction can be done via preparing/generate_keywords.py.

Tip

For the TCGA-{BRCA, STAD, THCA} cohorts, you can skip this step by using the provided CSV files in data

python ./preparing/generate_keywords.py

Stage 2. Align Vision-Language embeddings via Meta-Teacher

2a. Prepare data
The root_dir containing extracted features should follow the structure below:

root_dir/
├── uni/
├── gigapath/
├── virchow2/
└── medgemma/

2b. Train Stage-2

python train_stage2.py --config-name train-stage-2-configs

Stage 3. Language-Guided Adaptive Knowledge Distillation

python train_stage3.py --config-name train-stage-3-configs

Citation

Will be available soon

Acknowledgements

This implementation builds on ideas and components from several excellent open-source projects. We thank the authors of:

• GigaPath
• UNI
• Virchow2
• MedSigLIP
• TRIDENT(CLAM)

Please cite the corresponding papers when using their models, code, or training recipes.