Table 4 Effect of different augmentation schemes employed in this study. The models were first trained and tested without using train-time augmentations, in next experiments different combinations were used to assess the effect of image augmentations on the performance of proposed technique.
From: Melanoma segmentation using deep learning with test-time augmentations and conditional random fields
Augmentations | Model | Jaccard index |
|---|---|---|
No augmentations | UNet | 82.34 |
ResUNet | 79.22 | |
ResUNet++ | 82.87 | |
Rotation, vertical, and horizontal flip | UNet | 82.25 |
ResUNet | 79.28 | |
ResUNet++ | 83.01 | |
Rotation, vertical,horizontal flip, hue saturation value, RGB shift random brightness, and random contrast | UNet | 83.39 |
ResUNet | 80.35 | |
ResUNet++ | 83.44 | |
Rotation, vertical,horizontal flip, grayscale hue saturation value, RGB shift, random brightness random contrast, motion blur, and random gamma | UNet | 82.72 |
ResUNet | 79.81 | |
ResUNet++ | 83.10 |
