This repository contains my bachelor's thesis project for the Computer Science program at the Federal Institute of Brasília (IFB). The project focuses on developing a deep learning model to classify breast density in full-field digital mammograms (FFDM) according to the BI-RADS® standard. The goal is to assist radiologists in diagnosing breast cancer by automating density classification, a critical factor in early detection.
Breast density classification is a key component in mammographic analysis, as dense breast tissue can obscure tumors and increase cancer risk. This project explores state-of-the-art convolutional neural networks (CNNs) to classify breast density into four BI-RADS® categories (A, B, C, D). The work includes dataset integration, preprocessing, model training, and performance evaluation, with a focus on improving generalization and reproducibility.
Key contributions:
- Integration of multiple public mammography datasets.
- Novel multi-class training approach using binary classifiers.
- Extensive hyperparameter tuning and preprocessing pipelines.
- Open-source implementation for reproducibility.
- Clone the repository:
git clone https://github.com/loioladev/breast-density-classification.git
cd breast-density-classification -
Install the required dependencies using
uv syncor install each dependency found in the pyproject file manually usingpip. -
Download and prepare datasets to use in the project (see Datasets Preparation).
After downloading the datasets, you can process them using the following command:
uv run main.py convert <dataset> <path_to_dir>Where <dataset> is the dataset name (e.g., inbreast, cbis-ddsm, miniddsm) and <path_to_dir> is the path to the directory containing the dataset files. More information use the flag --help.
To train the model, first set the configurations in the config file and run the following command:
uv run main.py train -f <config_file>Where <config_file> is the path to the configuration file. More information use the flag --help.
For questions and collaborations, please contact me at matheusloiolapinto@gmail.com.