Finding Hidden Features Responsible For Machine Learning Failures In Breast Cancer Detection on Screening Mammography
Bachelor Project, May 2022.
To run this repository, a "/data" folder with the INbreast, CBIS-DDSM and potentially kau-bcmd data sets must be placed locally as siblings folders relative to the repository. Similarly, the end2end repository must be placed at the same level locally as this repository. This way, data and end2end models are fetched with relatived paths, i.e.: "../data/CBIS-DDSM/...".
/data_exploration: This folder exemplifies some data exploration of CBIS-DDSM and INbreast. "Heatmap generation.ipynb" contains code for patch classifier heatmaps, that utilizes functionality in the end2end Github.
/load_preprocess specifies functionality to load data sets from relative paths, and to transform data, i.e. from DICOM to PNG-16.
/model1: Here are the initial workings with the secondary research models, which were later discarded due to private data set.
/model2: Here are the workings with the end2end model Github. For example, test_cbis.py tests all the models on CBIS-DDSM test, and outputs instance-level posterior probabilities to later compute AUROCs. run_test_cbis.job is an example script ran on the ITU cluster, by setting up environment and calling test_cbis.py. Special prerequisites here are: Keras <= 2.0.8 and Tensorflow < 2.0.0.
/results_analysis contains functionality that extracts and computes results from the model outputs. For instance, get_results.py computes model-level AUROCS on both CBIS-DDSM and INbreast.