Related Experiment Videos
An Unsupervised Deep Learning Framework for Quantitative Breast Density Estimation from Mammograms
Khaldoon Alhusari1, Salam Dhou1
1Department of Computer Science and Engineering, American University of Sharjah, Sharjah P.O. Box 26666, United Arab Emirates.
Journal of Imaging
|July 27, 2026
Summary
This study introduces an unsupervised AI framework for accurate breast density estimation from mammograms, reducing subjectivity in cancer risk assessment. The method offers a practical tool for radiologists, improving early breast cancer detection and patient outcomes.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer is a leading diagnosis in women; early detection is crucial.
- Mammography assesses breast density, a key risk factor, but manual interpretation is subjective.
- Existing supervised methods inherit subjectivity from manual density labels.
Purpose of the Study:
- Develop an unsupervised framework for quantitative breast density estimation.
- Reduce subjectivity and inter-observer variability in mammographic density assessment.
- Provide a reliable tool to aid radiologists in clinical decision-making.
Main Methods:
- Proposed an unsupervised framework using a Convolutional Neural Network (CNN) for mammographic density segmentation.
- Implemented adaptive Region of Interest (ROI) extraction and a novel confidence metric.
- Utilized expert labels for post hoc calibration and a confidence-filtered majority voting scheme for classification.
Main Results:
- Achieved Silhouette scores > 0.92 for segmentation performance on DDSM and INbreast datasets.
- Demonstrated agreement with expert labels of 71.43% (DDSM) and 79.28% (INbreast).
- Unsupervised labeling showed effective clustering with average Silhouette scores of 0.57 (DDSM) and 0.50 (INbreast).
Conclusions:
- The unsupervised framework offers a practical, non-subjective method for quantitative breast density estimation.
- The approach has potential as a decision-support tool for radiologists.
- Further investigation is needed for clinical integration and validation.