A simple network module for pneumonia x-ray classification
- Overall Accuracy: 96.8%
- COVID-19 Recall/Precistion: 100%
A three-category classifier for pneumonia x-ray that distiguish non-pneumonia, normal pneumonia and COVID-19. We added CBAM (Convolutional Block Attention Module) before the first layer and after the last conbolutional layer of ResNet18 that significantly enhanced its ability.
Dataset 1: covid-chestxray-dataset(only COVID-19 data)
https://github.com/ieee8023/covid-chestxray-dataset
Dataset 2: CoronaHack-Chest X-Ray-Dataset(only normal pneumonia and non-pneumonia data)
https://www.kaggle.com/praveengovi/coronahack-chest-xraydataset
Integration pack: (contains all data above)
BaiduNetdisk: https://pan.baidu.com/s/11BwGCB1n2rQvHGc137_oLg
Password: dw7a
ResNet18 Classfication
We trained a ResNet18 model on the previous dataset by finetuning weights from Imagenet (Batch_size=24, Epoch=50, optim=Adam, learning_rate=0.001, criterion=CrossEntropy).
Results are as follows:
| ResNet18 | Precision | Recall | F1-score | Num |
|---|---|---|---|---|
| Normal | 0.94 | 0.88 | 0.91 | 151 |
| Pneumonia | 0.95 | 0.99 | 0.97 | 411 |
| COVID-19 | 1.00 | 0.70 | 0.82 | 23 |
CBAMResNet18 Classfication
We then added CBAM to the network with the same hyperparameter.
Results are as follows:
| CBAMResNet18 | Precision | Recall | F1-score | Num |
|---|---|---|---|---|
| Normal | 0.93 | 0.95 | 0.94 | 151 |
| Pneumonia | 0.98 | 0.97 | 0.98 | 411 |
| COVID-19 | 1.00 | 1.00 | 1.00 | 23 |
By using Grad-CAM, we can visualize the contribution of CBAM (in folder)