TaxoMIL: Taxonomy-Constrained Learning for Hierarchical Whole Slide Image Analysis [ECCV 2026]

Official implementation of TaxoMIL: Taxonomy-Constrained Learning for Hierarchical Whole Slide Image Analysis.

Framework

TaxoMIL consists of three main components:

1. A MIL backbone that aggregates patch-level WSI features into a slide-level representation.
2. A dual-branch conditioning module for coarse-level and fine-level diagnosis.
3. A dual-head text decoder that generates hierarchical diagnostic labels.

The framework is trained with text generation loss and taxonomy-guided auxiliary objectives.

Installation

git clone https://github.com/chaey0/TaxoMIL.git
cd TaxoMIL

conda create -n taxomil python=3.10 -y
conda activate taxomil

pip install -r requirements.txt

Data Preparation

TaxoMIL uses pre-extracted WSI feature bags as input.

Each CSV file should contain the following fields:

split,coarse_label,fine_label,feature_path

Example:

train,Benign,Usual Ductal Hyperplasia,/path/to/features/sample_001.pt
test,Malignant,Invasive Carcinoma,/path/to/features/sample_002.pt

Label definitions and taxonomy mappings are provided in:

config/

The expected feature format is a .pt file containing patch-level embeddings for each WSI.

Running Experiments

Training

Train the base model:

python main.py \
  --mode train \
  --data BRACS \
  --decoder_type GPT2

Train TaxoMIL with all taxonomy-guided objectives:

python main.py \
  --mode train \
  --data BRACS \
  --decoder_type GPT2 \
  --use_all_loss

Testing

python main.py \
  --mode test \
  --data BRACS \
  --ckpt_path checkpoints/BRACS/model.pth

Citation

If you find this repository useful, please consider citing our paper:

@inproceedings{taxomil2026,
  title     = {TaxoMIL: Taxonomy-Constrained Learning for Hierarchical Whole Slide Image Analysis},
  author    = {Lee, Chaeyeon and Nguyen Quoc, Khang and Song, Jinsol and Chong, Yosep and Yim, Kwangil and Kwak, Jin Tae},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}