Automatic assessment of mammographic density using a deep transfer learning method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammographic breast density is a significant risk factor for breast cancer.
• Radiologist assessments of density using visual analogue scales offer improved risk prediction.
• Accurate and consistent breast density assessment is crucial for cancer risk stratification.

Purpose of the Study:

• To develop automated deep learning models for mammographic breast density estimation.
• To train models using expert radiologist scores for accurate and consistent predictions.
• To improve breast cancer risk prediction through advanced computational methods.

Main Methods:

• Utilized a large dataset of nearly 160,000 mammograms with dual expert density scoring.
• Adapted two pre-trained deep neural networks to generate feature vectors.
• Employed linear and nonlinear regression for density prediction, comparing against a simulated optimal method.

Main Results:

• The deep learning model achieved a root mean squared error (RMSE) of 8.79 ± 0.21.
• Model predictions demonstrated consistency and performance levels similar to human expert assessments.
• Simulated optimal method yielded RMSEs between 7.98 and 8.90 for specific image views.

Conclusions:

• A deep learning framework utilizing transfer learning was successfully demonstrated for density estimation.
• The approach requires moderate computational resources and can be trained with limited data.
• This automated method holds potential for enhancing breast cancer risk assessment accuracy and efficiency.