Official implementation for MIDL 2023 paper: Inherently Interpretable Multi-Label Classification Using Class-Specific Counterfactuals Arxiv Paper | Attri-Net video on Youtube

conda create -n attrinet python=3.10
pip install -r requirements.txt

We perform evaluations with the following three Chest X-ray datasets.

CheXpert (https://stanfordmlgroup.github.io/competitions/chexpert/)

ChestX-ray8 (https://nihcc.app.box.com/v/ChestXray-NIHCC)

VinDr-CXR (https://vindr.ai/datasets/cxr)

We compared Attri-Net with the black model Resnet50 and an inherent interpretable model CoDA-Nets. We adapted the models slightly to our task settings (i.e the number of classes in the output is set to the number of diseases we trained on which is 5).

Resnet50 We use the PyTorch implementation of resnet50 in torchvision.models subpackage without using pretrained weights.

CoDA-Nets We use the official implementation of CoDA-Nets, and use the default parameters of large model "9L-L-CoDA-SQ-100000" defined to train on Chest X-ray images. We remove the WarmUpLR scheduler for more stable training for ChestX-ray8 and VinDr-CXR datasets.

If you use any of the code in this repository for your research, please cite as:

  @misc{sun2023inherently,
      title={Inherently Interpretable Multi-Label Classification Using Class-Specific Counterfactuals}, 
      author={Susu Sun and Stefano Woerner and Andreas Maier and Lisa M. Koch and Christian F. Baumgartner},
      year={2023},
      eprint={2303.00500},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}