Extended Data Fig. 1: mIoU localization performance of the saliency method pipeline on the test set using threshold values tuned on the validation set.

a, We first applied min–max normalization to the Grad-CAM saliency maps so that each value gets transformed into a decimal between 0 and 1. We then passed in a range of threshold values from 0.2 to 0.8 to create binary segmentations and plotted the mIoU score per pathology under each threshold on the validation set. The threshold that gives the max mIoU for each pathology is marked with an “X”. Pathologies are sorted alphabetically and shown in two plots for readability. b, Comparing mIoU localization performances of the saliency method pipeline on the test set (using the best thresholds tuned on the validation set) and the human benchmark. We found that the saliency method pipeline outperformed the human benchmark on two pathologies and underperformed the human benchmark on five pathologies. For the remaining three pathologies, the performance differences were not statistically significant. This finding is consistent with what we report in the manuscript using Otsu’s method.