BUSIS: A Benchmark for Breast Ultrasound Image Segmentation

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Healthcare

Background:

• Breast ultrasound (BUS) image segmentation is crucial for computer-aided diagnosis (CAD) systems.
• Existing BUS segmentation methods are often evaluated on small, private datasets, hindering objective performance comparison.
• A standardized benchmark is needed to assess and compare BUS segmentation algorithms.

Purpose of the Study:

• To establish a benchmark for B-mode breast ultrasound image segmentation.
• To objectively compare the performance of state-of-the-art segmentation methods.
• To identify effective segmentation strategies for clinical practice and research.

Main Methods:

• Collected and annotated 562 breast ultrasound images using standardized procedures with four radiologists.
• Compared the performance of 16 state-of-the-art segmentation algorithms, including deep learning and conventional approaches.
• Utilized a losses-based approach to evaluate the sensitivity of semi-automatic segmentation to user interactions.

Main Results:

• Deep learning-based segmentation methods achieved high Dice Similarity Coefficient (DSC) values (≥ 0.90).
• Deep learning approaches significantly outperformed conventional methods in BUS image segmentation.
• The benchmark provides objective performance metrics for various segmentation strategies.

Conclusions:

• Deep learning models represent the current state-of-the-art for breast ultrasound image segmentation.
• The developed benchmark facilitates objective comparison and evaluation of segmentation algorithms.
• Further research should focus on refining segmentation strategies for improved clinical application and theoretical understanding.