Meeting 06.08.20 - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki

What happened so far

Infrastructure / Tensorboard
- Add class based metrics to tensorboard for all metrics
- fixed masked metrics
Preprocessing
Architecures
- Masked Loss function
Model training
- base TF: -> equal to ours
- base pytorch: -> equal to ours
- pytorch chexpert leadership top4 implementation
  - 5 classes: Cardiomegaly, Edema, Consolidation, Atelectasis, Pleural Effusion
    - 320x320 without data augmentation: 2 Epochs: Auc : 0.828 0.899 0.888 0.839 0.915,Mean auc: 0.874
    - 256x256 3 Epochs: Auc : 0.770 0.926 0.900 0.826 0.929,Mean auc: 0.870
    - 320x320 3 Epochs: Auc : 0.795 0.929 0.905 0.810 0.931,Mean auc: 0.874
  - 12 classes: Enlarged Cardiomediastinum, Cardiomegaly, Lung Opacity, Lung Lesion, Edema, Consolidation, Pneumonia, Atelectasis, Pneumothorax, Pleural Effusion, Pleural Other, Fracture
    - 320x320: Auc : 0.878 0.501 0.826 0.918 0.464 0.929 0.916 0.635 0.828 0.864 0.932 0.974,Mean auc: 0.806
- chexpert dataset paper:
  - 5 classes: auc: 0.811, 0.840, 0.932, 0.929, 0.931
- ensemble paper:
  - 5 classes: auc: 0.745, 0.813, 0.882, 0.921, 0.930
- ours (trained on 63% of train data):
  - 12 classes (DenseNet121_Chexpert_BCE_E3_B32_C1_N12)
    - Mean auc: 0.723

Test auc_enlarged_cardiomediastinum:  0.6274955868721008
Test auc_cardiomegaly:  0.827772855758667
Test auc_lung_opacity:  0.70615553855896
Test auc_lung_lesion:  0.6908576488494873
Test auc_edema:  0.7949528694152832
Test auc_consolidation:  0.7065780758857727
Test auc_pneumonia:  0.6720331907272339
Test auc_atelectasis:  0.6417831778526306
Test auc_pneumothorax:  0.7809439301490784
Test auc_pleural_effusion:  0.8473460078239441
Test auc_pleural_other:  0.7237358093261719
Test auc_fracture:  0.6555337309837341

Other
- Visualizing Predictions
  - Grad Cam: Gradient-weighted Class Activation Mapping
  - Guided Grad Cam
  - CNN Fixations

Problems, questions & discussion points

server issues

Next steps

integrate conditional training into our framework (benchmark/experiments)
add label smoothing regularization
retrain with full dataset
normalize weights for wbce, cwbce
analyse leaderboard github code
add same transformations ( data augmentation)
maybe implement visualization pytorch
NO masked loss for WBCE
learning rate anpassen

Next meeting:

20.08.2020 14:00 Uhr