Capability and reliability of deep learning models to make density predictions on low-dose mammograms

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Breast density is a key factor in cancer risk assessment.
• Digital mammography and deep learning models are used for automated density estimation.
• Low-dose mammography offers potential benefits, especially for younger women, but its utility for density estimation needs evaluation.

Purpose of the Study:

• To evaluate the capacity and reliability of deep learning models for predicting breast density from low-dose mammograms.
• To assess the feasibility of using low-dose mammography for cancer risk estimates in younger women.
• To analyze factors influencing the accuracy of density prediction from low-dose mammograms.

Main Methods:

• Deep learning models were trained on both standard-dose and simulated low-dose mammograms.
• Model performance was tested on a dataset with paired standard- and low-dose images.
• The impact of factors like age, breast density, and dose ratio on prediction accuracy was analyzed, alongside methods to improve performance.

Main Results:

• Deep learning models demonstrated reliable breast density prediction from low-dose mammograms, comparable to standard-dose images.
• Breast area significantly impacts prediction accuracy; higher correlation (0.985) was observed for larger breasts versus smaller breasts (0.882).
• Averaging predictions across different views (CC-MLO) and multiple model trainings improved performance.

Conclusions:

• Low-dose mammography is a viable tool for generating breast density and cancer risk estimates comparable to standard-dose imaging.
• Averaging techniques can enhance the predictive performance of deep learning models for breast density.
• Model accuracy for density prediction is reduced in cases of denser and smaller breasts.