Table 4.
Architecture of expert 3
| Input Image | Layer | Layer type | Size | Output shape |
|---|---|---|---|---|
| Across x axis | 1 | Convolution + ReLu | Filters 12, kernel 3 × 3 | (12, 93, 66) |
| Across y axis | (12, 77, 66) | |||
| Across z axis | (12, 77, 93) | |||
| Across x axis | 2 | Max pooling | Kernel 2 × 2, stride 2 | (12, 46, 33) |
| Across y axis | (12, 38, 33) | |||
| Across z axis | (12, 38, 46) | |||
| Across x axis | 3 | Convolution + ReLu | Filters 12, kernel 3 × 3 | (12, 44, 31) |
| Across y axis | (12, 36, 31) | |||
| Across z axis | (12, 36, 44) | |||
| Across x axis | 4 | Max pooling | Kernel 2 × 2, stride 2 | (12, 22, 15) |
| Across y axis | (12, 18, 15) | |||
| Across z axis | (12, 18, 22) | |||
| Across x axis | 5 | Convolution + ReLu | Filters 12, kernel 3 × 3 | (12, 20, 13) |
| Across y axis | (12, 16, 13) | |||
| Across z axis | (12, 16, 20) | |||
| Across x axis | 6 | Max pooling | Kernel 2 × 2, stride 2 | (12, 10, 6) |
| Across y axis | (12, 8, 6) | |||
| Across z axis | (12, 8, 10) | |||
| Across x axis | 7 | Fully connected + ReLu dropout (rate 0.5) | 500 hidden units | 500 |
| Across y axis | 500 | |||
| Across z axis | 500 | |||
| Across x axis | 8 | Softmax | 2 | 2 |
| Across y axis | 2 | |||
| Across z axis | 2 |