Problem with pretrained model on MIMIC-CXR or MIMIC-CXR-JPG

[LiamLiu62]

Hi! Thank you so much for your amazing works.

I want to know the pretrained model "densenet121-res224-all" was trained on MIMIC-CXR (DICOM version) or MIMIC-CXR-JPG. I find that my results on my MIMIC-CXR-JPG test set are worse than the Benchmark shown.

Thanks.

[ieee8023]

It was trained on the JPG files. To confirm I verified the dataset does not use a dicom loader to read the files: https://github.com/mlmed/torchxrayvision/blob/73effd43b1f664ade43bb5043c8c27422663f89a/torchxrayvision/datasets.py#L1219

The script for generating the benchmark results is here: https://github.com/mlmed/torchxrayvision/blob/master/scripts/model_calibrate.py Maybe there is some difference you can see?