Mass Segmentation in Automated 3-D Breast Ultrasound Using Adaptive Region Growing and Supervised Edge-Based...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Automated 3-D breast ultrasound is a complementary tool to mammography for early breast cancer detection.
• Accurate mass segmentation is crucial for feature extraction and temporal comparisons in computer-aided detection (CAD) systems.
• Automated segmentation of breast masses is challenging due to variations in shape, size, and texture.

Purpose of the Study:

• To develop a novel computerized segmentation system for 3-D breast ultrasound masses using only a seed position as prior information.
• To improve the accuracy and reliability of mass segmentation for enhanced computer-aided detection and diagnosis.

Main Methods:

• A two-stage segmentation approach was developed, incorporating shape information from training masses.
• The first stage utilized an adaptive region growing algorithm with a similarity threshold determined by a Gaussian mixture model.
• The second stage employed a novel geometric edge-based deformable model, initialized with the results from the first stage.

Main Results:

• The proposed two-stage method achieved a mean Dice score of 0.74 ± 0.19 on a dataset of 50 masses (38 malignant, 12 benign).
• This performance significantly outperformed adaptive region growing (mean Dice 0.65 ± 0.2, p < 0.02) and distance regularized level set evolution (mean Dice 0.52 ± 0.27, p < 0.001).
• The supervised method demonstrated accurate mass segmentation results, validated by the Dice measure.

Conclusions:

• The developed computerized segmentation system provides accurate mass segmentation in 3-D breast ultrasound images.
• This method can be used for automated measurement of breast lesion volume changes over time.
• The segmentation results can be utilized to extract features for computer-aided detection or diagnosis systems, aiding in early breast cancer detection.