Fig. 2.

Binary cross-entropy loss and DSC on the training and validation sets during training of the left-hemithorax deep CNN [(a) without data augmentation and (b) with data augmentation] and of the right-hemithorax deep CNN [(c) without data augmentation and (d) with data augmentation]. Solid lines indicate results on the initial pairs of training/validation sets. Shaded areas indicate the range of the loss and DSC across all three pairs of training/validation sets used to assess variance in segmentation performance during training. The vertical dashed lines [at epoch 19 in (b) and epoch 12 in (d)] indicate the training epochs after which the deep CNNs trained on the initial training/validation sets were applied to the test sets.