A Two-Step Segmentation Method for Breast Ultrasound Masses Based on Multi-resolution Analysis.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Radiology

Background:

• Breast ultrasound is valuable for cancer detection but suffers from noise and low contrast, complicating mass segmentation.
• Accurate segmentation of breast masses is crucial for reliable computer-aided diagnosis and treatment planning.

Purpose of the Study:

• To propose a fully automated, two-stage approach for segmenting breast masses in ultrasound images.
• To evaluate the performance of AdaBoost and active contours algorithms in the second segmentation stage.

Main Methods:

• Initial segmentation using support vector machine or discriminant analysis with multiresolution pixel descriptors, non-linear diffusion, and curvature measures.
• Automated region of interest selection via heuristic rules.
• Refined segmentation using AdaBoost with scale-variant curvature and non-linear diffusion, or active contours initialized with first-stage results.

Main Results:

• Promising segmentation performance achieved by both AdaBoost and active contours.
• Normalized Dice similarity coefficients of 0.824 (AdaBoost) and 0.813 (active contours).
• High precision rates of 89.3% for both methods, with recall rates of 79.6% (AdaBoost) and 77.8% (active contours).

Conclusions:

• The proposed two-stage automated segmentation method effectively segments breast masses in ultrasound images.
• Both AdaBoost and active contours demonstrate robust performance, offering potential for improved breast cancer detection systems.