EDRN - Preprocessing Mammograms for AI

Standardizing Breast Screening Data for Deep Learning

This repository provides a comprehensive preprocessing pipeline for 2D mammogram images, ensuring standardization, consistency, and usability for deep learning applications in breast cancer screening. The pipeline converts raw DICOM images into AI-ready datasets, including structured metadata and statistical visualizations.

Overview

Deep learning models require well-structured and preprocessed datasets for reliable training and evaluation. This repository introduces a robust preprocessing framework to prepare mammogram data for AI-driven analysis. The dataset comprises three matched case-control studies (N=349, 318, 180 matched case-controls), with each patient having at least four DICOM images:

• L-CC (Left Craniocaudal)
• R-CC (Right Craniocaudal)
• L-MLO (Left Mediolateral Oblique)
• R-MLO (Right Mediolateral Oblique)

Additionally, clinical metadata—including BI-RADS classifications, study logs, and hormonal replacement therapy usage history—is integrated into structured JSON files for ease of use in deep learning models.

Key Features

✅ DICOM Processing

• Converts DICOM images to PNG format for better compatibility and storage efficiency.
• Resizes images to 1664×2048 pixels.
• Aligns images to a common coordinate space.
• Normalizes images to zero mean and unit variance.

✅ Data Augmentation

• Applies random horizontal/vertical flips, rotations (up to 20°), affine transformations, and color jitter to enhance model generalizability.

✅ Metadata Extraction & Organization

• Extracts DICOM headers and integrates them with clinical study logs, BI-RADS classifications, and hormonal history.
• Saves metadata in structured JSON files for streamlined analysis.

✅ Dataset Splitting

• Partitions data into training-validation (80%) and test (20%) sets.
• Further splits the training-validation set into 80%-20% partitions for robust model evaluation.

✅ Statistical Visualizations

• Provides insights into age distribution, image count, laterality, view type, density, BI-RADS classification, cancer prevalence, and implant presence.

✅ Reproducibility & Open-Source Availability

• The entire pipeline is available as a Jupyter Notebook, enabling easy integration into deep learning workflows.

Getting Started

1. Clone the Repository

git clone https://github.com/lab-rasool/EDRN_Breast_Data_Preprocess.git
cd EDRN_Breast_Data_Preprocess  

2. Install Dependencies

Ensure you have Python installed. You can install the required packages using:

pip install -r requirements.txt  

3. Run the Preprocessing Notebook

Open the Jupyter Notebook and execute the preprocessing steps:

jupyter notebook Process_mammograms.ipynb  

4. Output Files

• Processed PNG images (aligned, normalized, and augmented).
• JSON metadata files containing clinical and imaging information.
• Data splits (training-validation-test).
• Visualizations for exploratory analysis.

Contributing

Contributions are welcome! Feel free to fork the repository, submit issues, or create pull requests.

License

This project is released under the MIT License.

Acknowledgments

Special thanks to the research community advancing AI applications in breast cancer screening.