Keyword [ChestX-ray14] [Attention]
Guan Q, Huang Y, Zhong Z, et al. Diagnose like a radiologist: Attention guided convolutional neural network for thorax disease classification[J]. arXiv preprint arXiv:1801.09927, 2018.
1. Overview
1.1. Motivation
- existing methods use global image as input
- data limitation
- small localized areas
- poor alignment
In this paper, it proposes AG-CNN (Attention Guided CNN)
- three-branch. global, local and fusion
- learn a global CNN branch using global images
- crop local region based on global branch
- fuse global and local
1.2. Related Work
- JSRT dataset
- Shenzhen chest X-ray set
- Montgomery County chest X-ray set
- Indiana University Chest X-ray Colletion dataset
2. Architecture
2.1. AG-CNN
- output. [l1, l2, …, l15]. 14 disease + 1 No Finding
2.2. Algorithm
2.3. Attention Guided Mask Inference
- absolute value
- (b, k, h, w) – (b, 1, h, w)
- binary mask. threshhold=0.7
- crop based on [x_min, y_min, x_max, y_max]
2.4. Training Strategy
- fine-tune global branch
- fine-tune local branch, fix global branch
- fine-tune fusion branch, fix local and global branch
3. Experiments
3.1. Dataset
- randomly shuffle 70-10-20
3.2. Details
- 256x256, randomly crop 224x224
- random horizontal flipping
- normalize by ImageNet mean value
- test. center crop
3.3. Comparison
3.4. Others
3.4.1. Threshold
3.4.2. Training Strategy
3.4.3. Heat Map Analysis
