Breast segmentation in screening mammograms using multiscale analysis and self-organizing maps

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Anatomy

Background:

• Accurate segmentation of the breast region in mammograms is crucial for diagnostic interpretation.
• Previous unsupervised self-organizing map (SOM) methods showed promise but required improvement in detecting low-contrast areas near the skin line and optimizing computational efficiency.

Purpose of the Study:

• To enhance the detection of the breast region near the skin line in screening mammograms.
• To reduce the computational complexity of breast region segmentation.
• To improve the clinical utility of automated mammogram analysis.

Main Methods:

• An improved unsupervised self-organizing map (SOM) technique was developed, exploiting global image properties at multiple scales.
• A multi-step strategy was implemented, including preliminary segmentation and a focused analysis of ambiguous pixels around the skin line.
• The refined SOM was applied to classify pixels within a defined band around the skin line, reducing computational load.

Main Results:

• The improved technique demonstrated enhanced detection of the low-contrast breast region near the skin line.
• Segmentation performance remained consistent across different mammographic views and breast densities.
• Computational cost was reduced by up to 80% compared to the original method.

Conclusions:

• The enhanced SOM-based segmentation method offers improved accuracy for detecting the breast region near the skin line in mammograms.
• Significant reduction in computational complexity makes the technique more efficient for clinical application.
• This advancement contributes to more robust and efficient automated analysis of screening mammograms.