This code implements our proposed DeepMiner framework and applies it to the DDSM dataset.
Please see the training directory for instructions to train a 3-class ResNet-152. Alternatively, you can use our pretrained model (resnet152_3class). The instructions for downloading the pretrained model are here.
Please see the unit_visualization directory for instructions to generate unit visualizations for the trained 3-class ResNet-152.
Please use generate_unit_rankings.py to rank the units. Since the final layer has 2048 units, it would be infeasible to annotate all of them, so we will find the most frequently influential units and annotate only those. If you are using our pretrained model, you can simply run:
python generate_unit_rankings.pyThe script will write out unit rankings to the output directory. By default, we select the top 20 units for each of the 3 classes, for a total of 60 units to annotate.
See the server directory to start up the expert annotation web interface. The top 60 units as ranked using generate_unit_rankings.py will be displayed under the "alternative ranking" section on the home page. You can view the 46 unit annotations done by our experts in the notebook cleaned_unit_labels.ipynb.
See the deepminer_explanations.ipynb notebook for code used to generate the DeepMiner explanations shown in our paper.
We carried out an experiment to test whether the DeepMiner explanations enable an expert to better distinguish malignant cases from benign cases. All relevant code for that experiment is contained in explanations_benchmark.ipynb.