Chexpert Leader JF Healtchare - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki
JF Healthcare is number 4 on the Leadership board and has provided their code on github
We used their code to retrain from scratch and adjusted it to work for any given number of classes and not just the 5 of the Leaderboard test.
We retrained both for 5 classes and for 12 classes as well as with or without any augmentations, or other "tricks".
Using augmentation and other "tricks" means:
- Data augmentation with affine transformations
- Data transformations:
- Histogram Equalization
- Gaussian Blur
- Using batch weights
The following settings where the same for all trainings:
- epochs: 3
- batch size: 32
- image size: 256x256
- optimizer: Adam
- learning rate: 1e-4
- momentum: 0.9
- learning rate factor: 1.0
- optimizer: Adam
- loss function: BCE
- batch norm for fully conntected layer: True
- drop out for fully conntected layer: 0.0
The following table shows the AUCs of these four settings and the results the authors reported themselves in their paper:
| Pathologies | Paper | N5 Aug | N5 NoAug | N12 Aug | N12 NoAug |
|---|---|---|---|---|---|
| Enlarged Cardiomediastinum | - | - | - | 0.9184 | 0.9016 |
| Cardiomegaly | 0.8703 | 0.8401 | 0.8704 | 0.5015 | 0.6231 |
| Lung Opacity | - | - | - | 0.8101 | 0.7872 |
| Lung Lesion | - | - | - | 0.9020 | 0.9226 |
| Edema | 0.9436 | 0.9031 | 0.9448 | 0.5376 | 0.8592 |
| Consolidation | 0.9334 | 0.9094 | 0.9205 | 0.9127 | 0.9177 |
| Pneumonia | - | - | - | 0.8921 | 0.8580 |
| Atelectasis | 0.9029 | 0.8229 | 0.7947 | 0.8404 | 0.7792 |
| Pneumothorax | - | - | - | 0.7917 | 0.8265 |
| Pleural Effusion | 0.9166 | 0.9214 | 0.9258 | 0.8316 | 0.8857 |
| Pleural Other | - | - | - | 0.9203 | 0.9199 |
| Fracture | - | - | - | 0.9095 | 0.9497 |