Related Experiment Video
Updated: Jan 9, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.5K
Generative AI-Enhanced Microcalcification Detection in Full-Field Digital Mammography: Reducing False Positives with
Kyungsu Kim1, Manisha Bahl1,2,3,4,5, Adham Mahmoud Alkhadrawi1,2,3,4,5
1School of Transdisciplinary Innovations, Interdisciplinary Programs in Artificial Intelligence & Biomedical Engineering, Department of Biomedical Science, Seoul National University, Seoul, Republic of Korea.
Research Square
|December 8, 2025
Summary
Generative AI significantly enhances calcification detection in mammograms, improving accuracy beyond current segmentation methods. This advancement promises to refine screening and reduce diagnostic errors in radiology.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Generative AI's role in radiology, particularly for calcification detection in full-field digital mammograms (FFDM), is not well-defined compared to segmentation AI.
- Initial segmentation AI achieved high sensitivity (98.0%) but suffered from low specificity, indicated by a positive predictive value (PPV) of 3.2%.
Purpose of the Study:
- To develop and evaluate a generative AI model for refining calcification detection in FFDM.
- To improve specificity and reduce false positives in mammogram analysis while maintaining high sensitivity.
Main Methods:
- A generative AI model was developed, utilizing segmentation AI output as a structural prior.
- The generative AI transformed calcification-positive pixels into calcification-free pixels and generated a corrected result via subtraction.
- The model was trained on true calcification-free regions, categorizing densities within and around calcifications.
Main Results:
- The generative AI approach improved PPV 2.28-fold (from 3.2% to 7.3%), outperforming previous generative AI models by 146-fold.
- Sensitivity was maintained above 95% throughout the process.
- Patient-level detection errors were significantly reduced for small calcifications (5.17-fold), high-exterior density (4.20-fold), and low-interior density (2.89-fold).
Conclusions:
- Generative AI demonstrates significant radiological value, surpassing current state-of-the-art segmentation models for calcification detection.
- This generative AI approach has the potential to redefine screening accuracy in mammography.
- The study highlights the capability of generative AI to create high-fidelity virtual normal FFDM references for further research and development.
