Table 1 Architectures of final CNN with different input patch-sizes based on trial and error
From: Microaneurysm detection in fundus images using a two-step convolutional neural network
Layer | Operation | Input size | Detail | Berr, (p) |
|---|---|---|---|---|
Layer 1 | Input | \(3\times 101\times 101\) | – | – |
Layer 2 | Convolutional | \(16\times 101\times 101\) | \(7\times 7\) | – |
Layer 3 | Max pooling | \(16\times 50\times 50\) | \(2\times 2\) | 0.25 |
Layer 4 | Convolutional | \(16\times 48\times 48\) | \(5\times 5\) | – |
Layer 5 | Max pooling | \(16\times 24\times 24\) | \(2\times 2\) | – |
Layer 6 | Convolutional | \(16\times 22\times 22\) | \(3\times 3\) | – |
Layer 7 | Max pooling | \(16\times 11\times 11\) | \(2\times 2\) | 0.25 |
Layer 8 | Convolutional | \(16\times 10\times 10\) | \(2\times 2\) | – |
Layer 9 | Max pooling | \(16\times 5\times 5\) | \(2\times 2\) | – |
Layer 10 | Convolutional | \(16\times 4\times 4\) | \(2\times 2\) | – |
Layer 11 | Max pooling | \(16\times 2\times 2\) | \(2\times 2\) | – |
Layer 12 | Fully connected | 100 | \(1\times 1\) | – |
Layer 13 | Fully connected | 2 | \(1\times 1\) | – |