Self-Supervised Deep Metric Learning for Prototypical Zero-shot Lesion Retrieval in Placenta Whole-Slide Images
Official repository for Self-Supervised Deep Metric Learning for Prototypical Zero-shot Lesion Retrieval in Placenta Whole-Slide Image.
Open access link to the article.
This repository contains the training pipeline presented in the paper, the evaluation code for CAMELYON16 and a demo notebook.
Model weights and a sample support set for CAMELYON16 can be downloaded from this repository.
You can install most requirements from requirements.txt.
pip install -r requirements.txt
You will also need to install ASAP and the multiresolutionimageinterface library to handle CAMELYON16 annotations.
An example configuration file is available in configs/local_train.conf. We advise you to read through the lib/parser.py file to see the different training parameters available. The training WSIs should be listed in a .json file as shown in ./camelyon_normal_split.json. When ready, start training with:
python train.py -c configs/local_train.conf
The patch- and slide-level zero-shot evaluation pipeline can be run from the eval.py file. The model_dict variable can be modified to evaluate other model/transforms. The slides to be used for prototype definition should be listed in a .json file as shown in ./camelyon_normal_split.json. To run patch-level evaluation, run:
python eval.py --model protossdml --name protossdml_camelyon16 --level patch
The demo.ipynb notebook contains an example of heatmap generation for a test WSI from CAMELYON16 using a support set generated during the zero-shot, simulated low-data regime evaluation.
If you find this repository useful, please consider giving a star ⭐ and cite the original article!
@article{protossdml_2025,
title = {Self-supervised deep metric learning for prototypical zero-shot lesion retrieval in placenta whole-slide images},
volume = {196},
issn = {0010-4825},
url = {https://www.sciencedirect.com/science/article/pii/S0010482525009850},
doi = {https://doi.org/10.1016/j.compbiomed.2025.110634},
journal = {Computers in Biology and Medicine},
author = {Faure, Gaspar and Soglio, Dorothée Dal and Patey, Natalie and Oligny, Luc and Girard, Sylvie and Séoud, Lama},
year = {2025},
pages = {110634},
}