Learning Causal Alignment For Reliable Disease Diagnosis (ICLR 2025)

Summary

If you happen to use or modify this code, please remember to cite our paper:

@inproceedings{
liu2025learning,
title={Learning Causal Alignment for Reliable Disease Diagnosis},
author={Mingzhou Liu and Ching-Wen Lee and Xinwei Sun and Xueqing Yu and Yu QIAO and Yizhou Wang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=ozZG5FXuTV}
}

Table with Results

Methodology	Precision of CAM(LIDC)	Precision of CAM(DDSM)	Classification Accuracy(LIDC)	Classification Accuracy(DDSM)
Ross et al. (2017)	0.034 (0.06)	0.084 (0.11)	0.656 (0.00)	0.559 (0.05)
Zhang et al. (2018)	0.068 (0.11)	0.110 (0.13)	0.381 (0.03)	0.581 (0.00)
Brendel & Bethge	0.048 (0.04)	0.090 (0.04)	0.358 (0.00)	0.592 (0.00)
Rieger et al. (2020)	0.041 (0.05)	0.232 (0.17)	0.343 (0.00)	0.586 (0.01)
Chang et al. (2021)	0.074 (0.03)	0.119 (0.07)	0.503 (0.08)	0.496 (0.08)
Oracle classifier	1.000 (0.00)	1.000 (0.00)	0.789 (0.00)	0.726 (0.01)
Ours	0.751 (0.03)	0.805 (0.06)	0.722 (0.00)	0.656 (0.00)

Installation

To install this repo, first create your virtual environment in Python3.8, for example this way:

python3.8 -m venv my_venv

And then install the requirements as follows:

pip install -r requirements.txt

Notes:

Please make sure your Torchopt version==0.7.0

Please make sure your Python version is consistent with the Pytorch, Torch Geometric, and CUDA versions you are going to install.

Running the experiments

sh .\DDSM\run.sh

sh .\LIDC\run.sh

Data Preprocess

Data Format in cache_path

• Image (\*_img.pt)
  • Shape: (batch_size, m, n)
  • Stores the mass images.
• Latent Code (\*_z.pt)
  • Shape: (batch_size, channel, w, h)
  • Stores the image's latent representation.
• Node ID (\*_nid.pt) or Jit ID (\*_jitid.pt)
  • Shape: (batch_size,)
  • Stores the mass node IDs.
• Labels (\*_label.pt)
  • Shape: (batch_size, 6)
  • Stores the mass attributes labels.
• Mask (\*_mask.pt)
  • Shape: (batch_size, m, n)
  • Stores the binary mask for each image (1 for mass region, 0 for background).

ccce_file

Stores mass attributes ccce score csv(cite from Conditional Counterfactual Causal Effect for Individual Attribution)

Code for CCCE https://github.com/LLily0703/CCCE

DDSM Data Preprocess code

• Raw data preprocess

  python DDSM\preprocess\test_db.py(cite from https://github.com/jbustospelegri/breast_cancer_diagnosis)

  python DDSM\preprocess\jit.py

  python DDSM\add_crossmask_manually.py

• auto encoder get latent code

  get DDSM image latent code by auto encoder (Call python3 train.py -c autoencoder_vanilla.json -id ${runid} -seed ${seed} -d ${device})

LIDC Data Preprocess code

• Raw data preprocess

  follow README.md in preprocess directory

• auto encoder get latent code

  get DDSM image latent code by auto encoder (Call python3 train.py -c autoencoder_vanilla.json -id ${runid} -seed ${seed} -d ${device})