(A) Structural similarity index measures (SSIMs) under
cascading randomization of modules on InceptionV3 for the SIIM-ACR
Pneumothorax Segmentation dataset and RSNA Pneumonia Detection Challenge
dataset. Note that the colored dotted lines correspond to the
randomization baseline for each saliency map, which were generated by
the average SSIMs of 50 randomly chosen pairs of saliency maps
pertaining to the fully trained model; a saliency model successfully
reaches degradation if it goes below its corresponding randomization
baseline. (B) Example image from RSNA pneumonia detection
dataset to visualize saliency map degradation from cascading
randomization. “Logits” refers to the logit layer (final
layer) of the InceptionV3 model, and layer blocks 1 through 10 refer to
blocks mixed 1 through mixed 10 in the original InceptionV3
architecture. ACR = American College of Radiology, GBP = guided
backpropagation, GCAM = gradient-weighted class activation mapping,
GGCAM = guided GCAM, GRAD = gradient explanation, IG = integrated
gradients, Orig = original, RSNA = Radiological Society of North
America, SG = Smoothgrad, SIG = smooth IG, SIIM = Society for Imaging
Informatics in Medicine.