Table 2 Receptive field (RF, in pixels), accuracy, and AUC from the test fold for different types of filter sizes, number of filters, depth of the network (resolution steps), using class weighting or binary labels (for U-net), and patch size (for SW-net)
From: Fully convolutional architecture vs sliding-window CNN for corneal endothelium cell segmentation
Method | RF | Acc. | AUC |
|---|---|---|---|
SW-net | |||
Patch 64pix, 32 filters of 3x3 | 61 | 95.82 | 0.9938 |
Default, Fully Connected Layer | 61 | 94.97 | 0.9916 |
Patch 96pix, 32 filters of 3 ×3 | 61 | 94.03 | 0.9888 |
Patch 96pix, 32 filters of 4 ×4 | 91 | 95.39 | 0.9931 |
U-net | |||
32 filters of 3 ×3, 4 steps | 61 | 97.55 | 0.9949 |
32 filters of 3 ×3, 5 steps | 125 | 97.62 | 0.9955 |
32 filters of 4x4, 4 steps | 91 | 97.65 | 0.9958 |
32 filters of 5 ×5, 4 steps | 121 | 97.46 | 0.9954 |
32 filters of 4 ×4, 3 steps | 43 | 97.48 | 0.9951 |
32 filters of 4 ×4, 5 steps | 187 | 96.92 | 0.9939 |
16 filters of 4 ×4, 4 steps | 91 | 97.32 | 0.9951 |
64 filters of 4 ×4, 4 steps | 91 | 97.61 | 0.9956 |
Default, weighted class | 91 | 96.65 | 0.9958 |
Default, binary labels | 91 | 93.92 | 0.99 19 |