Code and trained models for an external validation of saliency maps as error flags for chest radiograph classifiers

Code and trained models accompanying the paper "Can saliency maps flag chest radiograph classifier errors at a new hospital? An external validation study" by Aashish Chapagai, Angad Singh Gurtatta and Erik Cambria.

The study tests whether saliency maps from a DenseNet-121 and a ViT-B/16 chest radiograph classifier, trained on NIH ChestX-ray14, can detect the classifiers' errors at the training hospital and at two external hospitals (CheXpert and VinDr-CXR), beyond the models' own confidence.

This record contains:
- CXR-Saliency-External-Validation-main.zip: a snapshot of the GitHub repository, with the seven Jupyter notebooks, paper_figures.py, the environment files and the aggregate results reported in the paper
- cnn_base.keras and vit_base.keras: the trained DenseNet-121 and ViT-B/16 models (Keras 3)
- thresholds_macro.npy, thresholds_micro.npy, thresholds_hamming.npy and their vit_ equivalents: decision thresholds tuned on NIH validation data
- disease_classes.npy: the order of the 14 NIH findings
- test_probabilities.npy and vit_test_probabilities.npy: predicted probabilities on the NIH test set

No radiographs and no per-radiograph outputs from CheXpert or VinDr-CXR are included. All datasets are available from their providers under their own terms, as described in the README. The models were trained only on NIH ChestX-ray14, provided by the NIH Clinical Center.