Related Experiment Video
Updated: Oct 25, 2025
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
A Deep Learning Approach to Re-create Raw Full-Field Digital Mammograms for Breast Density and Texture Analysis.
Hai Shu1, Tingyu Chiang1, Peng Wei1
1Departments of Biostatistics (H.S., P.W., K.A.D.), Diagnostic Radiology (T.C., M.D.L., E.O.C., A.S., T.W.M., L.Q.C.S., J.W.T.L., O.O.W.), and Clinical Cancer Prevention (J.B.D., S.M.H.), The University of Texas MD Anderson Cancer Center, 1515 Holcombe Blvd, Houston, TX 77030; and Department of Biostatistics, School of Global Public Health, New York University, New York, NY (H.S.).
This study developed a deep learning method to reconstruct raw digital mammograms from processed ones, achieving high accuracy in image quality and breast density assessment.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Healthcare
- Computational Pathology
Background:
- Digital mammography generates processed images for routine screening.
- Raw, unprocessed mammograms contain valuable data but are rarely stored.
- Reconstructing raw mammograms could enhance diagnostic capabilities.
Purpose of the Study:
- To create a computational method for reconstructing raw digital mammograms from processed images.
- To assess the accuracy and reliability of the reconstructed raw mammograms.
Main Methods:
- A retrospective study utilized 3713 mammograms from 884 women.
- A deep learning model, U-Net convolutional network with kernel regression, was developed.
- Image quality, breast density, and 29 texture features were used for comparison.
Main Results:
- The deep learning approach accurately recreated raw mammograms with high similarity (SSIM=0.986).
- Reconstructed images showed strong correlation in breast density (r=0.946) and agreement in grade (κ=0.875).
- 23 out of 29 texture features showed satisfactory correlation, with others complemented by processed images.
Conclusions:
- The deep learning method effectively reconstructs raw mammograms from processed images.
- The approach demonstrates strong agreement in image metrics, breast density, and texture features.
- This technique holds potential for improving mammographic data analysis and interpretation.
