A deep learning method for classifying mammographic breast density categories.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Breast Cancer Risk Assessment
• Digital Mammography Image Analysis

Background:

• Mammographic breast density is a key breast cancer risk factor, assessed using BI-RADS categories.
• Radiologists face challenges in consistently differentiating 'scattered density' and 'heterogeneously dense' categories.
• Accurate breast density assessment is crucial for effective breast cancer screening and patient notification.

Purpose of the Study:

• To develop and evaluate a deep learning-based classifier for distinguishing between 'scattered density' and 'heterogeneously dense' BI-RADS categories.
• To provide a potential computerized tool to assist radiologists in consistent breast density classification.
• To improve the accuracy and reliability of breast density assessment in clinical workflows.

Main Methods:

• A convolutional neural network (CNN) model was trained on a dataset of 22,000 digital mammogram images.
• Both training from scratch and transfer learning approaches were evaluated for classification performance.
• Performance was measured using Area Under the Curve (AUC) on ROC curves, including analysis on a refined dataset.

Main Results:

• Training from scratch achieved an AUC of 0.9421, with accuracy increasing with more training data.
• Transfer learning with fine-tuning achieved an AUC of 0.9265 with only 500 images.
• After data refinement, AUCs increased to 0.9882 (from scratch) and 0.9857 (transfer learning), indicating high accuracy.

Conclusions:

• The deep learning approach demonstrated high accuracy in classifying difficult breast density categories.
• The developed classifier can potentially enhance the consistency of breast density assessment by radiologists.
• This technology may improve patient notification and risk stratification in breast cancer screening programs.