FIGURE 4.

Impact of data augmentations on attention heatmaps and segmentation results. The heatmaps calculated using Grad‐CAM highlight the attention regions. The four convolutional neural networks (CNNs) had the same U‐Net structure; however, trained by different datasets (A: baseline, B: baseline + noise injection [NI], C: baseline + random overlay [RO], D: baseline + NI + RO). DEC4, decoder block next to the bottom block; ENC4, encoder block next to the bottom block; GT, ground truth; pCTV, projected clinical‐target‐volume; XF, X‐ray fluoroscopic image