1 Introduction
2 Related work
3 Methods
3.1 Bag-of-visual-words
3.2 Convolutional bag-of-visual-words
3.3 Procedure
Layer | Output size |
---|---|
Input (dropout 30%) | 128 |
FC-ReLU (dropout 50%) | 500 |
FC-Softmax | 100 |
Layer | Output size |
---|---|
Input (dropout 30%) | 128 |
FC-ReLU (dropout 50%) | 150 |
FC-Softmax | 25 |
Layer | Filter size, stride | Output size |
---|---|---|
Input | - | 64 × 64 × 1 |
Conv-ReLU | 5 ×5, 1 | 60 × 60 × 16 |
Max-pool | 2 ×2, 2 | 30 × 30 × 16 |
Conv-ReLU | 3 ×3, 1 | 28 × 28 × 16 |
Max-pool | 2 ×2, 2 | 14 × 14 × 16 |
Conv-ReLU | 3 ×3, 1 | 12 × 12 × 16 |
Max-pool | 2 ×2, 2 | 6 × 6 × 16 |
Conv-ReLU | 3 ×3, 1 | 4 × 4 × 16 |
Max-pool | 2 ×2, 2 | 2 × 2 × 16 |
Flatten | - | 64 |
FC-Softmax | - | 32 |
4 Evaluation
4.1 Datasets
Dataset | Normal | Pathological |
---|---|---|
DR1 | 595 | 482 |
DR2 | 337 | 98 |
Messidor | 546 | 654 |
4.2 Results
Method | AUC |
---|---|
Sparse SURF | 93% ± 1% |
Dense CNN | 91% |
Method | AUC |
---|---|
Pires et al. (2014) [6] | 94% |
Dense SURF | 95% ± 1% |
Sparse SURF
|
97% ± 1%
|
Dense CNN
|
97%
|