Determination of mammographic breast density using a deep convolutional neural network

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• High breast density is a significant risk factor for breast cancer.
• Accurate breast density classification is crucial for risk assessment and screening decisions.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (dCNN) for automated breast density classification.
• To classify breast density based on mammographic appearance according to the American College of Radiology Breast Imaging Reporting and Data System (ACR BI-RADS) Atlas.

Main Methods:

• A dCNN model with 11 convolutional and 3 fully connected layers was trained on 20,578 mammography views from 5221 patients.
• The dataset was augmented and sorted by ACR density from radiological reports.
• The model was tested against radiological reports and a consensus of two human readers on separate datasets.

Main Results:

• The dCNN achieved high accuracy on validation datasets (MLO: 90.9%, CC: 90.1%).
• On test datasets, algorithm agreement with radiological reports was 71.7% (MLO) and 71.0% (CC), improving to 88.6% (MLO) and 89.9% (CC) for dense vs. fatty differentiation.
• Agreement with human readers was high (MLO-model: 92.2%, CC-model: 87.4%), reaching 99% (MLO) and 96% (CC) for differentiating dense from fatty breasts.

Conclusions:

• The developed dCNN accurately classifies breast density according to the ACR BI-RADS system.
• This automated technique provides standardized and observer-independent breast density evaluation.
• Accurate, standardized classification can reduce misclassification, enhance breast cancer risk assessment, and guide decisions on supplemental screening like ultrasound.