Skip to main content
. 2021 Jun 24;26(9):1125–1137. doi: 10.1177/24725552211023214

Figure 4.

Figure 4.

More training data from a relevant domain improves accuracy. (A) Pixel-wise F1 score on the A549 cell line (y axis) for models trained plainly (solid line), with label smoothing (dashed line) or with data augmentation (dotted line) for U-Net, U-Net++, Deeplabv3+, Tiramisu, and PPU-Net (panels, colors) for an increasing number of training images (x axis). (B) Pixel-wise F1 score for the U-Net model on the A549 cell line (y axis) for an increasing number of training images (x axis), fine-tuning on the target domain (dashed line) or source and target domains (solid line) and testing on the source domain (red line) or target domain (blue line). Source domain of six of seven cell lines in the seven cell line data set; target domain the seventh cell line. (C) As in B, but using the source domain of the LNCaP data set and target domain of the seven cell line data set. (D) Pixel-wise F1 scores (y axis) for all models (colors) for an increasing number of focal planes (x axis) used as input during training.