To-do

[muellerdo]

• [x] Implement directory interface in AUCMEDI
• [x] Implement inference augmenting in AUCMEDI
• [x] Implement augmenting predictions in baseline as well for comparison with normal baseline prediction
• [x] Run baseline again
• [x] Evaluate baseline
• [x] Compute baseline predictions for evaluation set with some model (upload and see score on online evaluation set)
• [x] Implement up-sampling approach for rare label combinations
• [x] Run classifier 5-fold CV with selected architectures
• [x] Implement ensembler & ensembler Training
• [x] Run ensembler training
• [x] Implement prediction
• [x] Test out single model for each class ML algorithms
• [x] Implement sample weights for RF label approach
• [x] Implement disease risk binary model and train it
• [x] Implement focal loss to AUCMEDI
• [x] Implement multi-label class weights in AUCMEDI
• [x] Implement ensemble learning for disease risk
• [x] Implement single architecture inference for Focal Loss EfficientNetB4 and upload
• [x] Implement focal loss usage to all models

Further:

• [x] Test out Single Model Learning for each class Ensemble with Random Forest
• [x] Implement disease risk to ensembling
• [x] Project structuring
• [x] Try out longer transfer learning epochs (20?) with lower learning rate (E-07?)
• [x] Increase epochs to 300
• [x] Try out focal loss https://github.com/sushanttripathy/Keras_loss_functions/blob/master/focal_loss.py https://stackoverflow.com/questions/48485870/multi-label-classification-with-class-weights-in-keras

Final:

• [x] Implement Evaluation
• [x] Re-run all classifiers and detector trainings
• [x] Re-run ensemble training
• [x] Re-run inferences
• [x] Re-run ensemble
• [x] Upload
• [x] Run Evaluation on final data
• [x] Write Manuscript
• [x] Revise Manuscript
• [x] Update README
• [x] Upload Data to Zendo
• [x] Make this repo public

Upload:

• [x] Simple LR approach
• [x] Simple RF approach
• [x] Augmenting LR approach
• [x] Augmenting RF approach

Idea:

• Run predictions on training dataset (detector & label)
• Collect all predictions (detector & label)
• Train a LR/RF Model for each class

[muellerdo]

Utilize inference augmenting mean

[muellerdo]

Alternative Strategy:

• Resize image via subfunctions to (512, 512)

• Set resize to None in DataGenerators

• Apply Cropping with (224,244) to image augmentation for training

• Apply cropping to inference as well

[muellerdo]

Extra Model for disease risk classification

[muellerdo]

-> Rank 25 overall and team rank 18