Geometry-Based Pectoral Muscle Segmentation From MLO Mammogram Views.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Computer-aided diagnosis (CADx) systems are crucial for early breast cancer detection.
• Accurate segmentation of the breast region, including pectoral muscle removal from MLO mammograms, is essential for CADx performance.
• Existing pectoral muscle segmentation methods fail to address all breast density types.

Purpose of the Study:

• To develop a novel, effective method for segmenting pectoral muscles across all breast density classes in MLO mammograms.
• To improve the accuracy and reliability of CADx systems by addressing limitations in pectoral muscle segmentation.

Main Methods:

• A new method combining geometric rules and a region growing algorithm was developed.
• The method supports segmentation of all pectoral muscle types: normal, convex, concave, and combinatorial.
• Segmentation accuracy was evaluated on 872 MLO images from three public datasets, categorized by four tissue density classes.

Main Results:

• Achieved an average Jaccard index of 0.972 ± 0.003 and Dice similarity coefficient of 0.985 ± 0.001.
• Demonstrated a mean Hausdorff distance below 5 mm for all datasets.
• Attained an average acceptable segmentation rate of approximately 95%, outperforming state-of-the-art methods, especially for dense breasts.

Conclusions:

• The proposed method effectively segments pectoral muscles in MLO mammograms across all breast densities.
• This advancement has the potential to enhance the accuracy of computer-aided diagnosis systems for breast cancer.
• The method shows excellent performance, even on challenging cases with extremely dense breasts.