MIL-RNN nature-medicine-2019

This repository provides training and testing scripts for the article Campanella et al. 2019.

Directory organisation

• MIL-Train : Training MIL
• MIL-Test : Inference MIL
• RNN_train : Train RNN
• RNN_Test : Test RNN
• CheckSet.py: Open an check all images in {Train,Val}.json
• CheckQuick.py: Open an check all images in {Train,Val}.json
• Untitled.ipynb : Split Typical/Atypical To rename
• Untitled1.ipynb : Experience data augmentation To rename
• Res*.ipynb : Jupyter notebook model summarizing the results
• {Train*,Val*}.json: Files need for torch DataLoader

Dataset

{ 'Slides':['TNE1095' , 'TNE1411', ...],

'Tiles' : [[TNE1095/Tiles1095_x_y.jpg, ...] [TNE1411/TNE1411_x_y.jpg, ...] ],

'Targets': [0, 0, 1 ... ] }

Main modification from the original folder

• MIL:
  • DataLoader
  • Data augmentation:
    • Removed the normizalition
    • Add color augmentation
  • Output modification:
    • Save learning rate
    • Calculation of Training error, FPR, FNR
  • Training with learning rate schedule
  • Possibility to load a graph
• RNN:
  • Add Attention layer

Experiment 1: Typical/Atypical

Try to locate atypical area.

Dataset:

• ~1M Tiles 512x512px Vahadane color normalization Typical/Atypical (see Untitled.ipynb).

Experience:

A - ResNet34 - Batch size 32

• Epoch 25
• Huge instability
• From ImageNet

B - Resnet34 - Batch size 64

• Epoch 30
• Huge instability
• From exp A
• • RNN

Experiment 2: Tumoral/No Tumoral

See xlsx file to get a summary

Results organisation:

• ModelName:
  • Checkpoint_best.pth
  • convergence.csv
  • prediction.csv // From MIL_test.py
  • probability.csv // From MIL_test.py
  • *.ipynb // results summary
  • BestTiles:
    • SampleName_{0,1}
      • Normal
        • Tiles_{}.jpg
      • Tumour
        • TIles_{}.jpg
• ResMapVal
• ResMapTrain