Leveraging uncertainty information from deep neural networks for disease detection

Code and models for Christian Leibig, Vaneeda Allken, Murat Seckin Ayhan, Philipp Berens, Siegfried Wahl (Scientific Reports, 2017) / (preprint, 2016), developed at the ZEISS Vision Science Lab in collaboration with the Berenslab @ University of Tuebingen.

Getting started

If you want to use the Bayesian CNNs for detecting diabetic retinopathy with uncertainty have a look at disease-detection/example.ipynb.

To get things running you need a machine with a NVIDIA GPU and install nvidia-docker. The docker image can be built as follows: Clone the repository and cd into the folder disease-detection and execute:

docker build -t uncertain-ai-diagnostics -f docker/Dockerfile .

Next, start a Docker container:

nvidia-docker run -it -p 8888:8888 uncertain-ai-diagnostics

This will fire up a jupyter notebook server and tell you the URL you have to point your browser to in order to play around with the example notebook.

Contact

leibig.christian@gmail.com