Table 1 U-Net architecture used for the DCGMM in our study
Layer | Details | Â | Layer | Details |
|---|---|---|---|---|
Input | Â | Â | Output | Â |
↓ |  |  | ↑ |  |
↓ |  |  | conv9_3 | 1x1x32; ReLU |
↓ |  |  | conv9_2 | 3x3x64; ReLU |
conv1_1 | 3x3x32; ReLU | Â | conv9_1 | 3x3x64; ReLU |
conv1_2 | 3x3x32; ReLU | → | concat4 | concatenate upsample4 with conv1_2 |
↓ |  |  | ↑ |  |
pool1 | 2x2 max pool stride 2 |  | ↑ |  |
↓ |  |  | upsample4 | 2x2 upsample of conv8 |
conv2_1 | 3x3x64; ReLU | Â | conv8 | 3x3x32; ReLU |
conv2_2 | 3x3x64;ReLU | → | concat3 | concatenate upsample3 with conv2_2 |
↓ |  |  | ↑ |  |
pool2 | 2x2 max pool stride 2 |  | ↑ |  |
↓ |  |  | upsample3 | 2x2 upsample of conv7 |
conv3_1 | 3x3x128; ReLU | Â | conv7 | 3x3x64; ReLU |
conv3_2 | 1x1x128; ReLU |  | ↑ |  |
conv3_3 | 1x1x128; ReLU | → | concat2 | concatenate upsample2 with conv3_3 |
↓ |  |  | ↑ |  |
pool3 | 2x2 max pool stride 2 |  | ↑ |  |
conv4_1 | 3x3x256; ReLU | Â | upsample2 | 2x2 upsample of conv6 |
conv4_2 | 3x3x256; ReLU | Â | conv6 | 3x3x128; ReLU |
conv4_3 | 1x1x256;ReLU | → | concat1 | concatenate upsample1 with conv4_3 |
↓ |  |  | ↑ |  |
pool4 | 2x2 max pool stride 2 |  | ↑ |  |
conv5_1 | 3x3x256; ReLU | Â | upsample1 | 2x2 upsample of conv5_3 |
conv5_2 | 3x3x256; ReLU |  | ↑ |  |
conv5_3 | 1x1x256; ReLU | → | ↑ |  |