Related Experiment Video
Updated: Aug 23, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.0K
Deep Learning Models for Automated Assessment of Breast Density Using Multiple Mammographic Image Types
Bastien Rigaud1, Olena O Weaver2,3, Jennifer B Dennison4
1Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX 77030, USA.
Cancers
|October 27, 2022
Summary
This study introduces a deep learning model, EfficientNetB0, for automated breast density assessment using multiple mammogram types. The model shows substantial agreement, outperforming human experts and commercial software for reliable breast density classification.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Automated breast density assessment is crucial for breast cancer risk stratification.
- Existing deep learning models primarily use single mammographic views.
- Clinical data integration with multi-modal mammography for density assessment remains underexplored.
Purpose of the Study:
- To investigate pre-trained EfficientNetB0 deep learning models for automated breast density assessment.
- To evaluate the impact of using multiple mammographic types (FFDM, DBT, synthesized 2D) with and without clinical information.
- To compare the model's performance against inter-observer variability and commercial software.
Main Methods:
- Retrospective analysis of 120,000 mammograms from 5032 women across multiple screening examinations.
- Optimization of pre-trained EfficientNetB0 deep learning models with and without clinical history.
- Evaluation using BI-RADS (4-category) and binary (dense/non-dense) breast density classifications.
- Performance comparison using Fleiss' Kappa scores against inter-observer and Volpara software agreement.
Main Results:
- Human expert agreement (Fleiss' Kappa) ranged from 0.31-0.69, indicating significant uncertainty.
- Commercial software (Volpara) showed fair to moderate agreement (0.33-0.54).
- Proposed EfficientNetB0 models achieved moderate to substantial agreement (0.61-0.75), outperforming human and commercial software variability.
- Optimal breast density estimation was achieved using FFDM and DBT images without additional clinical information.
Conclusions:
- Pre-trained EfficientNetB0 models offer a reliable and versatile method for automated breast density assessment.
- The model demonstrates superior performance compared to human observers and commercial software.
- FFDM and DBT images, commonly archived in clinical practice, yield the best results for automated density assessment using this DL approach.
Keywords:
deep learning (DL)digital breast tomosynthesis (DBT)full-field digital mammograms (FFDM)synthesized 2D images
