breast_cancer_classifier

Project Introduction: Breast Cancer Classifier

Overview

The breast cancer classifier project is a vital initiative designed to aid radiologists in the early detection of breast cancer through improved accuracy in mammography screenings. Rooted in the research paper titled "Deep Neural Networks Improve Radiologists' Performance in Breast Cancer Screening," this project harnesses the power of deep learning to enhance diagnostic processes.

Purpose

The primary aim of the project is to develop a system that offers accurate predictions of breast cancer presence by analyzing mammography images. It utilizes two primary models for this purpose: one that uses only images (image-only model) and another that combines images with heatmaps (image-and-heatmaps model). These models are crucial tools that provide probabilities of benign or malignant findings for each breast.

Input and Output

• Input Images: The system analyzes two standard mammographic views per breast: CC (Cranio-Caudal) and MLO (Mediolateral oblique), requiring 16-bit PNG images.
• Input Heatmaps: Heatmaps, representing benign and malignant categories, are generated to match the size of the mammography images.
• Output: The models generate four predictions: left_benign, right_benign, left_malignant, and right_malignant, which indicate the likelihood of benign or malignant conditions in each breast.

How it Works

The project is implemented in PyTorch and operates through a comprehensive pipeline consisting of four stages:

1. Cropping: Mammograms are cropped to focus on the breast region, removing unnecessary background.
2. Center Calculation: Optimal window centers are calculated for accurate imaging.
3. Heatmap Generation: Heatmaps are created to assist in diagnosis.
4. Classification: The models process the images (with or without heatmaps) to predict cancer presence.

For those interested in using or modifying the system, an understanding of the necessary technical prerequisites is essential, including proficiency in Python and familiarity with libraries such as PyTorch, SciPy, and OpenCV.

Running the Model

• View-Wise Model: The more comprehensive model analyzes all four mammographic views per breast exam.
• Image-Wise Model: A simpler version that works with single images, suitable for cases where complete exams are unavailable. Although it performs slightly less accurately than the view-wise model, it's beneficial for specific use cases.

To execute the model, users can utilize the provided shell script run.sh, or execute individual Python scripts for more control over the processing and output stages. Optimal performance is achieved using a GPU, but the system can revert to CPU use if necessary.

Data Handling and Preprocessing

The input data must be preprocessed to meet the model’s requirements, involving conversion from DICOM to PNG format to maintain high pixel granularity. The preprocessing steps ensure that images are standardized and ready for analysis, significantly enhancing the model's ability to accurately predict cancer presence.

Legal and Ethical Considerations

This project is shared under the GNU AGPLv3 license, emphasizing ethical use and collaboration in the development and application of diagnostic tools.

Conclusion

The breast cancer classifier project represents significant progress in medical imaging and AI, offering tools that promise to assist radiologists in making more accurate diagnoses. However, it remains a research-focused project not yet sanctioned for clinical use. For those interested in exploring recent advancements, collaboration opportunities exist, with contact information provided for further inquiry.