Breast density characterization using texton distributions.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Breast density is a significant risk factor for breast cancer.
• Current breast density classification methods (BI-RADS) suffer from observer variability.
• Accurate breast density assessment is crucial for risk stratification.

Purpose of the Study:

• To develop a fully automated breast density classification system.
• To introduce a novel texture classification method using textons and TDSM.
• To improve the accuracy and consistency of breast density assessment.

Main Methods:

• Representing mammograms using textons (clustered filter responses).
• Characterizing breast density patterns via the texton spatial dependence matrix (TDSM).
• Utilizing chi-square distance for classification based on normalized TDSM matrices.

Main Results:

• The automated TDSM method achieved over 82% classification accuracy on the Oxford Mammogram Database.
• The TDSM approach demonstrated superior performance compared to simple texton histograms.
• Texton spatial dependencies enhance breast density classification accuracy.

Conclusions:

• The proposed automated TDSM method offers a robust approach for breast density classification.
• This technique has the potential to reduce inter-observer variability in mammogram interpretation.
• Accurate automated breast density classification can aid in personalized breast cancer risk assessment.