Keyword [ChestX-ray14]
Zhou B, Li Y, Wang J. A weakly supervised adaptive densenet for classifying thoracic diseases and identifying abnormalities[J]. arXiv preprint arXiv:1807.01257, 2018.
1. Overview
In this paper,
- adaptive DenseNet
- bridging layer
- WSL pooling strucuture
2. Architecture
2.1. Adaptive DenseNet
- remove avg pooling from the third transition layer
- dilated all kernels in the fourth dense block
- output 14x14
2.2. Bridging Layer
- transform by 1x1 Conv
- (b, 1664, h, w) – (b, MxC, h, w)
2.3. WSL Pooling
2.3.1. Class-Wise Pooling
- C (b, M, h, w) – C (b, 1, h, w) – (b, C, h, w)
2.3.2. Spatial-Wise Pooling
- (b, C, h, w) – (b, C, 1, 1)
- during training, randomly select from top-k m
- duting testing, both top-k and bottom-k
3. Experiments
3.1. Dataset
- same as ChestX-ray14 split
- 256x256, randomly crop 224x224
- normalize by ImageNet
- test. center crop
3.2. Details
- weighted BCE
- in each DenseBlock, add BN and dropout 0.1
- M = 14
- training. k=10
- testing. k+ = k- = 15
- α = 1
- heatmap threshold. 0.8 for Cardiomegaly, 0.9 for others
3.3. Comparison
