Official implementation of the paper: Fusion of Multi-scale Heterogeneous Pathology Foundation Models for Whole Slide Image Analysis. Code for biomarker prediction is available at this moment. Complete usage will be released upon paper accepatance.
Ubuntu 20.04, CUDA version 12.0.
Whole slide image (WSI) and gene expression dataset: The WSIs used in this work are from TCGA (https://portal.gdc.cancer.gov/), open access to all.
python==3.10
torch==2.6.0
torchvision==0.21.0
huggingface-hub==0.30.2
openslide-python==1.4.2
trident
mvlearn
Complete usage will be released upon paper accepatance.
python train.py --wsi_train_feat_dir './training_emds' \
--model_output_path './results' \
--wsi_valid_feat_dir './validation_emds' \
--label_path './labels' \
--teacher_emds_path_1 './teacher_1_emds' \
--teacher_emds_path_2 './teacher_2_emds' \
--teacher_emds_path_3 './teacher_3_emds' \
--num_classes 2 \
--epochs 60 \
--batch-size 64 \
--lr 0.0001 \
--seed 2025
We sinceresly thank following work with their open-sourced code.
Wentao,Tang et al. Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology. CVPR 2024. https://doi.ieeecomputersociety.org/10.1109/CVPR52733.2024.01078
Andrew Z et al. Accelerating Data Processing and Benchmarking of AI Models for Pathology. arXiv 2025. https://arxiv.org/pdf/2502.06750