Figure 3
From: Impact of image compression on deep learning-based mammogram classification
Saliency map results for a left craniocaudal view of a “malignant” case. Subfigure (a): the original image with a CR of one. Subfigure (b): the radiologist’s annotation on the original mammogram (i.e., ground truth). Subfigures (c)∼(f): saliency maps from models trained on images with the same compression ratio on images with CRs of 1, 100, 5 K, and 11 K. The areas marked in red are areas significant for prediction for each model.
