Keyword [ChestX-ray14]
Wang H, Xia Y. Chestnet: A deep neural network for classification of thoracic diseases on chest radiography[J]. arXiv preprint arXiv:1807.03058, 2018.
1. Overview
In this paper, it incorporates the attention mechanism into DNN
- classification branch
- attention branch. Grad-CAM
- experiments on Chest X-ray 14 dataset
1.1. Related Work
- different pooling strategy
- statistical label dependencies
- CheXNet. dense connection + BN
1.2. Dataset
- Chest X-ray 14 dataset with official patient-wise split (80%/20%)
- 10% among 80% as validation
1.3. Model
- ResNet-152. remove softmax layer; replace last FC; Sigmoid
1.3.1. Attention Branch
- choose output of penultimate residual module
- first 3 Conv: 1x1, 3x3, 1x1
- A~. output of the third CNN
- A-. map for each class c
- α_ck. computed by using the gradient propagation of Grad-CAM
- Then normalize each element in A-.
- use normalized map A as input to last 3 CNN (14, 1x1; 512, 1x1; 1, 14x14)
1.3.2. Training
- 224x224 image
- no data augmentation
- threshold 0.5
1.4. Experiments
1.4.1. With or Without Attention
