Meeting 28.05.20 - TobiasSchmidtDE/DeepL-MedicalImaging GitHub Wiki

What happened so far

Datasets
- Dataset exploration for chestX-Ray14 NIH
- Updated exploration for chexpert
- MIMIC-CXR Dataset (in progress)
Infrastructure
- integrate remote access into docker build file
- fix training on gpu & added logging script for gpu
- set up tensorboard & tensorboard callbacks
- Bug fixes in test and model evaluation pipeline
Architecures
- Implement SOTA Guendel (done, tested implemention on single batch)
Preprocessing
- Minor refactoring and bug fixes generator class
- Utils for test, train, validation split based on patient id
Other
- Explore differnt loss functions
- More metrics
  - AUC
  - Mean Binary Accuracy
  - Precisin & Recall
  - F2 Score
- "Fix" zero values in classifciation report
- Standardized experiments, benchmarks and model creation
Model training
- Retrain models for chexpert and chestxray14
  - Chexpert: Resnet, Densenet (done), Inceptionnet(running), SOTA Guendel (queued)
  - Chestxray14: Resnet, Densenet, Inceptionnet, SOTA Guendel (queued)

merging datasets (postponed)
- Chexpert and chestX-Ray14 NIH
- rerun experiments with merged datasets
use weighted mean binary cross-entropy loss function
figure out why perfomance is low
- train on all (most) layers, not only last dense layer
improve training performance (GPU utilisation)
- reduce batch size to 64
fix "metrics duplications" in tensorboard
get MIMIC-CXR dataset
better preprocessing, try out stuff
- smart cropping
- ...
improve documentation of benchmark, and experiment class
further research into chexpert leadership board papers