Mammographic Density from BUS

---

Deep Learning Predicts Mammographic Breast Density in Clinical Breast Ultrasound Images

Github repository containing all relevant code for journal submission

This repository is designed to provide implementation of the training, evaluation, and validation for our deep learning model which predicts the ACR Breast Imaging and Reporting Data System (BI-RADS) breast composition for mammography categories from breast ultrasound (BUS).

For model evaluation, use BUSClean (possibly tuned on your dataset) to prepare input BUS scans. Doppler and elastography scans should be excluded, scans should be split if dual-view, scans with calipers and text removed or cropped out from the image. Invalid images and images collected as part of a biposy should be removed.

Results

Installation and system requirements

• Tested on Ubuntu 20.04.6 LTS
• Python version: 3.9.16
• To install dependencies, run:

python setup.py install

Demo

• Demo scripts are provided in the outermost folder.
• Model checkpoint files (as used for results reported in the paper) can be downloaded from here.
• A sample dataset is provided purely to demonstrate model functionality, the dataset is not representative of the complete dataset used to train/evaluate the models in the manuscript. Note: the sample data provided have not been cleaned/cropped as expected by the model.
• Demo files are described below:
  • write_graylevels.py pulls out image histograms from included datasets and saves them as pickle files.
  • graylevel_method.py does grid searches and saves predictions from an MLP, random forest and logistic regression models from the image histograms written above.
  • generate_predictions.py generates predictions from trained deep learning models.
  • train.py provides model training code, given optimal hyperparameters found during Optuna runs.