This is the source code for the MICCAI 2018 Paper, Exploring Uncertainty Measures in Deep Networks for Multiple Sclerosis Lesion Detection and Segmentation (Nair et al.), of which I am the first author.
The network architecture is a heavily modified U-Net (Ronneberger et al.), developed in Tensorflow. The network is augmented to provide the following 4 different uncertainty measures as an output.
- Mutual Information (Gal et al.)
- Entropy (Gal et al.)
- MC Sample Variance (Leibig et al.)
- Predicted Variance (Kendall and Gal)
Details about the network architecture, and the equations for the uncertainty measures can be found in the paper here: https://arxiv.org/abs/1808.01200
The dataset used for this project comes from a large, proprietary, multi-site, multi-scanner, clinical MS dataset. As such, to use this code you will have to modify the dataprovider to be specific to your dataset.
Training:
- pip install -r requirements.txt
- python bunet_launcher.py -o ./path_to_output/ -c bunet/configs/train_bunet.json
Author: Tanya Nair