histopathology_cancer_detection

Project for HPI SS2020 for the Deep Learning lecture

Authors:

• Nicolas Alder (Master Data Engineering, Hasso Plattner Institute)
• Eric Fischer (Master Data Engineering, Hasso Plattner Institute)
• Erik Langenhan (Master IT Systems Engineering, Hasso Plattner Institute)
• Nataniel Müller (Master Digital Health, Hasso Plattner Institute)
• Christian Warmuth (Master Data Engineering, Hasso Plattner Institute)
• Simon Witzke (Master Data Engineering, Hasso Plattner Institute)

Assignment

Histopathologic cancer detection on hematoxylin and eosin (H&E)-stained lymph node sections slides following the Kaggle challenge.

Running and Training on Colab

How to install and let it run can be viewed in this demo.

Our training colab notebooks can be viewed here:

• VGGNet11
• VGGNet19
• ResNet18
• ResNet152
• DenseNet121
• DenseNet201

Setup and Training on Server

Setup

pip3 install -r requirements.txt

Training

python3 -W ignore train.py -trnl "path to train split" -tstl "path to test split" -f "path to images" -o "path to output" -m densenet121 -n "Neptune Text"

Testing

python3 test.py -tstl "path to test split" -f "path to images" -a densenet121 -p uuid-model.pkl -o uuid-opt.pkl  -hist uuid-history.json

Experiment Documentation

For documentation and comprehensibility, we stored all our runs in a Neptune.ai project here.

Code References

For normalization, we used and modified the python implementation of the normalization method proposed by Macenko et al. under the following licence (a copy also visible in our repo here). This python implementation is used in the file helper_scripts/normalization.py.