The application produces segmentations using saliency method heatmaps and human annotations, assessing the localization performance of these segmentations based on the methodology outlined in the paper titled "Benchmarking Saliency Methods for Chest X-ray Interpretation."

Certain conditions, like effusions and cardiomegaly, tend to appear consistently in similar locations on frontal view chest X-rays (CXRs), whereas the locations of others, such as lesions and opacities, can differ across CXRs. The exploration of how saliency methods affect user trust and efficacy is limited.

Based on the paper: https://www.nature.com/articles/s42256-022-00536-x.

Details on model weights used are on Github (https://github.com/rajpurkarlab/cheXlocalize/tree/master?tab=readme-ov-file#weights).

You can run data using your own heatmaps/annotations/segmentations or on the Chexlocalize dataset (https://stanfordaimi.azurewebsites.net/datasets/23c56a0d-15de-405b-87c8-99c30138950c). 

Here is the link to sample data: https://github.com/rajpurkarlab/cheXlocalize/tree/master/sample