This is an implementation of a basic Faster R-CNN model for breast tumor detection on the ultra-high resolution Duke BCS-DBT (Breast Cancer Screening - Digital Breast Tomosynthesis) dataset, to be used as a "baseline" model for the DBTex breast lesion detection challenge and benchmark.
We provide notebooks from pre-processing and dataset preparation to training and testing stages. We offer two version of the Faster-RCNN implementation.
We build our implemenation mainly based on MMDection. The installation of MMDetection can be found here.
The raw (DICOM) image data and annotation/label tables can be downloaded from the BCS-DBT page on The Cancer Imaging Archive. After downloading the tables, put them in ./data+_csv.
We supply the notebook preprocess.ipynb to do data preprocessing, which can transfer the raw dicom images into png image slices and create the corresponding json files.
For model training, use the notebook train&evaluate.ipynb.
Trained models are saved in /work_dirs/breast. If you want to evaluate our pretrained checkpoints, you can set the cfg.work_dirs to the pretrained model.
For testing, we offer the notebook generate_test_table.ipynb to create predictions formatted according to the DBTex guidelines.