将图像特征建模成不同解剖结构的Graph，用于下游的多标签分类和报告生成。
- Knowledge Graph
- Nodes from the same organ are connected
-
- Node embedding
- DenseNet-121 pretrained on CheXpert
- Attention:
- 1*1 conv+softmax
- each class node uses the corresponding channel as spatial attention
- Nodes from the same organ are connected
- Training Loss
- Node Level: binary classification of each node
- Graph Level: multi-label classification of 14 diseases
