Extended Data Fig. 9: Hit/miss: Coefficients from regressions on model assurance.
From: Benchmarking saliency methods for chest X-ray interpretation
Statistical analysis to determine whether there was any correlation between the model’s confidence in its prediction and saliency method pipeline performance using hit/miss. We used the true positive slice of the dataset (CXRs with both the most representative point identified by the saliency method/human benchmark and also the ground-truth segmentation). Since every heat map contains a maximally activated point (the pixel with the highest value) regardless of model probability output, using the full dataset has limited value since false positives are due to metric set up and are not associated with model probability.
