Computer-aided assessment of breast density: comparison of supervised deep learning and feature-based statistical...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast density is a significant risk factor for cancer.
• Accurate breast density assessment is crucial for risk stratification.
• Current methods for density estimation can be subjective and time-consuming.

Purpose of the Study:

• To develop and evaluate a supervised deep learning approach for automated percentage density (PD) estimation on digital mammograms (DMs).
• To compare the performance of the deep convolutional neural network (DCNN) approach against a traditional feature-based statistical learning method.

Main Methods:

• Digital mammograms underwent log-transformation and multi-resolution preprocessing.
• A deep convolutional neural network (DCNN) was trained using a domain adaptation resampling method to estimate a probability map of breast density (PMD).
• Percentage density (PD) was calculated from the PMD, and the DCNN approach was compared to a feature-based method using extracted gray level, texture, and morphological features.

Main Results:

• The DCNN approach achieved a Dice's coefficient (DC) of 0.79 ± 0.13 and Pearson's correlation (r) of 0.97 during cross-validation, outperforming the feature-based method (DC = 0.72 ± 0.18, r = 0.85).
• On an independent test set, the DCNN achieved DC = 0.76 ± 0.09 and r = 0.94, while the feature-based method achieved DC = 0.62 ± 0.21 and r = 0.75.
• The DCNN approach demonstrated significantly better performance and robustness compared to the feature-based learning approach.

Conclusions:

• The developed DCNN approach provides a robust and accurate method for automated percentage density estimation on digital mammograms.
• This automated approach has the potential to improve the efficiency and consistency of density reporting.
• The DCNN's performance suggests its utility in model-based breast cancer risk prediction.