Shifting the emphasis: Deep-learning models trained on original radiographs (left) focus on the image labels; models trained on radiographs with the label covered (right) emphasize anatomic features ...