Learning to Harmonize Cross-vendor X-ray Images by Non-linear Image Dynamics Correction
Run download-embed.py to download the data required by the experiment, and run prepare-embed.py to process the EMBED dataset.
Additionally, download the RSNA pneumonia detection dataset from this [link](RSNA Pneumonia Detection Challenge (2018) | RSNA), and run prepare-rsna.py to process the RSNA dataset.
After that, run prepare-dataset.py to pre-process the data, you can specify which dataset you will use by changing dataset_name.
For evaluation purposes only, skip this section and simply follow the instructions in the "test" section. You can download our pre-trained weights from this link and unzip to the "checkpoint" folder.
Run train-patho.py to train the pathology classifier (breast density classification or phenomina detection).
For evaluation purposes only, skip this section and simply follow the instructions in the "test" section.
Specify the checkpoint folder of your pathology classifier obtained from the previous step, as well as the target scanner you want to align to the reference one. Below is an example:
machine_config = ["Clearview", "embed-patho-full-20250317_143344"]
and then run train-enhance.py to train GDCE.
Test without GDCE: Replace the downstream task model checkpoint folder. Below is an example:
classifier_dir = 'embed-patho-full-20250317_143344'
Run test-enhance.py to get the 5-fold cross-validation results.
Test with GDCE: Replace the default GDCE and downstream task model checkpoint folders. Below is an example:
machine_config = ["Clearview", "embed-clearview-20250406_142325", "embed-patho-full-20250317_143344"]
Run test-enhance.py to get the 5-fold cross-validation results.