Meeting 22.07.20 - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki

BigBlueButton Meeting Room

What happened so far

• Infrastructure

  • Add learning rate graph to tensorboard and validation notebook (reduce LR on plateau)

• Preprocessing

  • Improved GPU util (~60% -> ~90%+) by preprocessing full dataset in advance

• Architecures

  • SotA Ensemble: conditional training, see [wiki entry](CheXpert SotA)
  • Masked Loss function
  • Removed "no finding" and "support device" classes

• Model training

  • retrained Densenet and InceptioNet on different combinations of:
    • CWCBE, WBCE and Masked CWBCE Loss
    • 14, 13, 12 classes (all, without nofinding, without nofinding and support devices)
  • Sota Ensemble (5 Classes)

• Other

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

• Model improvements:

  • DenseNet121_Chexpert_CWBCE_E10_B32_C1_N14_Masked_NoNan
    • Best CWBCE Validation AUC with non zero precision/recall
    • best test results (all metrics)
  • DenseNet121_Chexpert_CWBCE_E10_B32_C1_N13
  • InceptionV3_Chexpert_CWBCE_E10_B32_C1_N14_Masked_NoNan
  • InceptionV3_Chexpert_CWBCE_E10_B32_C1_N13
  • InceptionV3_Chexpert_CWBCE_E10_B32_C1_N12
    • Best minority class recall

Problems, questions & discussion points

• disease hierarchy
• minory class improvements
• over all model performance
  • still nowhere near 85%+ AUC
• radio exam

Next steps

• integrate conditional training into our framework (benchmark/experiments)
• implement visualization
• masked loss for WBCE
• fix masked models validation / classification report
• add metrics for individual pathologies
• add label smoothing regularization
• rerun basic experiment with pytorch

Next meeting:

06.08.2020 14:00 Uhr