Self-supervised diffusion pretraining for thoracic pathology detection offers insights into confounders
See requirements.txt
pip install -r requirements.txt
We provide checkpoints for DiffChest model trained on a joint collection of CXRs, namely, MIMIC-CXR (USA), CheXpert (USA), and PadChest (Spain). In addition, we also offer latent statistics and the classifier model finetuned on high-quality PadChest subset which was annotated by physicians.
Checkpoints ought to be put into a separate directory checkpoints.
Download the checkpoints and put them into checkpoints directory. It should look like this:
checkpoints/
- padchest_autoenc
- last.ckpt # DiffChest checkpoint
- padchest_autoenc_cls
- last.ckpt # Finetuned logistic regression classifier
Please register and download those publicly available CXR datasets:
After downloading and preprocessing those images, you need to create a LMDB dataset for model training.
We provide scripts for training & evaluate DiffChest on the following datasets: MIMIC-CXR, CheXpert, and PadChest.
Note: Most experiment requires at least 3x 3090s during training the DPM models while requiring 1x 2080Ti during training the accompanying classification head.
PadChest: 256
We only trained the DiffChest due to the high computation cost. This requires 3x 3090s.
python run_padchest.py
After the previous stage, a classifier (for manipulation) can be trained using:
python run_padchest_cls.py
We provide a testing script for evaluating the classification performance of DiffChest:
python test_padchest_cls.py
To generate visual explanations described in our paper, run:
python manipulate.py
If you use DiffChest, we would appreciate your references to our paper.
Please open new issue threads specifying the issue with the codebase or report issues directly to than@ukaachen.de.
The source code for the site is licensed under the MIT license, which you can find in the LICENSE file.
The official implementation of diffae model: https://github.com/phizaz/diffae