Related Experiment Video
Updated: Aug 9, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.0K
Unsupervised anomaly detection with generative adversarial networks in mammography
Seungju Park1, Kyung Hwa Lee2, Beomseok Ko3
1Department of Biomedical Engineering, College of Health Sciences, Korea University, Seoul, Republic of Korea.
Scientific Reports
|February 22, 2023
Summary
This study introduces a new AI method using generative models to create synthetic mammograms and detect breast cancer. This approach shows promise for improving early breast cancer screening, especially when labeled data is scarce.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer screening relies heavily on mammography.
- Deep learning advancements are enhancing mammography analysis.
- Supervised learning for breast cancer detection faces challenges due to limited labeled abnormal data.
Purpose of the Study:
- To develop an unsupervised method for breast cancer detection using synthetic mammograms.
- To leverage generative models for creating high-quality synthetic mammographic images.
- To assess the efficacy of anomaly detection for identifying breast cancer on mammograms.
Main Methods:
- Utilized StyleGAN2, a generative network, to synthesize realistic mammographic images.
- Trained the generative model exclusively on normal mammograms.
- Implemented an anomaly detection approach for breast cancer classification on a separate dataset.
Main Results:
- The generative model produced synthetic images with fidelity comparable to real mammograms.
- The anomaly detection method demonstrated high sensitivity in identifying breast cancer.
- The approach successfully differentiated between normal and cancer-positive mammograms.
Conclusions:
- The proposed generative and anomaly detection model shows potential for effective breast cancer screening.
- This unsupervised method can overcome limitations of supervised learning in data-scarce clinical settings.
- The technique offers a viable tool to aid in the early detection of breast cancer.
