Keyword [ChestX-ray14]
Yan C, Yao J, Li R, et al. Weakly supervised deep learning for thoracic disease classification and localization on chest x-rays[C]//Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. ACM, 2018: 103-110.
1. Overview
1.1. Motivation
- existing methods do not treat different diseases separately
In this paper, it exploits DenseNet-121 as backbone, and equip with
- SE block
- multi-map transfer
- max-min pooling
1.2. Related Work
- JSRT dataset
- BSE-JSRT dataset
- Indiana chast X-ray
- MIMIC-CXR dataset. lateral views are available
- AGCNN
2. Architecture
2.1. SE Block
model the interdependency between channels
- (b, c, h, w) –avg– (b, c, 1, 1)
- (b, c, 1, 1) –FC+FC– (b, c, 1, 1)
1. r. reduction rate
- (b, c, 1, 1) * (b, c, h, w)
2.2. Multi-Map Layer
- (b, c, h, w) -M 1x1_Conv- (b, Mc, h, w). (M: the number of class)
- (b, Mc, h, w) –Classwise Pooling– (b, c, h, w)
2.3. Max-Min Pooling
- (b, c, h, w) – (b, c)
- z^c. c-th pooled feature map
- k+ = k- = 1
- α = 0.7
3. Experiments
3.1. Dataset
- ChestX-ray14. official 70-10-20
- 512 x 512, 3RGB, normalize
3.2. Details
- BCE loss
- Adam with 0.0001, *0.1 when 5 time
- batch size 16
- random crop 448x448 (4 corner + 1 center)
- horizontal flipping
- reimplement CheXNet
- not using BBox