Keyword [ChestX-ray14] [PLCO]
Guendel S, Grbic S, Georgescu B, et al. Learning to recognize abnormalities in chest x-rays with location-aware dense networks[C]//Iberoamerican Congress on Pattern Recognition. Springer, Cham, 2018: 757-765.
1. Overview
1.1. Motivation
- most methods report performance based on random image splitting, ignore the high probability of the same patient appearing in both training and test test
- most methods fails to explicitly incorporate the spatial information of abnormalities or utilize the high resolution images
In this paper, it proposes location aware Dense Networks (DNetLoc)
- incorporate spatial information and high resolution for classification
- provide new reference patient-wise splits for ChestX-ray14 and PLCO
1.2. Related Work
- Wang. proposes ChestX-ray14 dataset
- Chexnet. slightlymodify DenseNet
- Yao. DenseNet + LSTM (randomly split dataset)
- Guan. attention guided CNN (randomly split dataset)
1.3. Dataset
1.3.1. ChestX-Ray14
- 1024x1024, 8 bits gray-scale
- 112,120 images
1.3.2. PLCO
- 2500x2100, 16 bits gray-scale
- choose 12 most prevalent pathology labels, among which 5 labels contains spatial information
Across both data sets, there are 6 labels which share the same name, but not combing them
1.4. Method
- loss function
- leverage high-resolution images (2 Conv with 3 kernel, 3x3 size, stride 2). 1024x1024 as input
- 35 lable where 21 from PLCO
1.5. Experiments
