Automatic Breast Ultrasound Image Segmentation: A Survey (1704.01472v2)

Abstract: Breast cancer is one of the leading causes of cancer death among women worldwide. In clinical routine, automatic breast ultrasound (BUS) image segmentation is very challenging and essential for cancer diagnosis and treatment planning. Many BUS segmentation approaches have been studied in the last two decades, and have been proved to be effective on private datasets. Currently, the advancement of BUS image segmentation seems to meet its bottleneck. The improvement of the performance is increasingly challenging, and only few new approaches were published in the last several years. It is the time to look at the field by reviewing previous approaches comprehensively and to investigate the future directions. In this paper, we study the basic ideas, theories, pros and cons of the approaches, group them into categories, and extensively review each category in depth by discussing the principles, application issues, and advantages/disadvantages.

• The paper presents a comprehensive review of BUS segmentation techniques crucial for early breast cancer diagnosis.
• It categorizes methods into graph-based, deformable, learning-based, and classical approaches while analyzing their benefits and limitations.
• The survey advocates for standardized benchmarks and improved deep learning models to advance automatic tumor delineation.

Automatic Breast Ultrasound Image Segmentation: A Survey

The survey titled "Automatic Breast Ultrasound Image Segmentation: A Survey" provides a comprehensive overview of the methodologies employed in the domain of breast ultrasound (BUS) image segmentation, addressing the exigent need for improving cancer diagnosis through early detection. Recognizing the prevalence of breast cancer as a critical global health issue, ultrasound imaging is acknowledged for its significance due to its cost-effectiveness, non-invasiveness, and applicability in diverse medical settings, particularly in resource-constrained regions.

Technical Discussion

The paper categorizes existing BUS segmentation techniques into six primary methodologies: graph-based approaches, deformable models, learning-based approaches, thresholding, region growing, and watershed. Each category of techniques is critically examined regarding their principles, merits, limitations, and application in clinical practices.

1. Graph-Based Approaches: The surveyed graph-based methods, which include MRF-MAP and Graph Cuts, leverage graphical models for efficiently representing segmentation tasks. These approaches suffer primarily from achieving only local optima (MRF-MAP) and face challenges such as the "shrink" problem in Graph Cuts-based segmentation. The extension of graph models through high-order connections is suggested to improve the representation of smoothness terms.
2. Deformable Models (DMs): These include parametric (PDMs) and geometric (GDMs) models, critical for capturing anatomical variability in BUS images. While PDMs offer speed and intuitive control, their sensitivity to initialization limits their robustness. GDMs overcome topology adaptation limitations inherent in PDMs but are computationally demanding.
3. Learning-Based Approaches: Machine learning methods, both supervised and unsupervised, are recognized for their ability to integrate diverse feature spaces for improving segmentation. Despite their performance in preliminary segmentation stages, refinements are often necessary to achieve precise tumor delineation.
4. Classical Methods: Techniques such as thresholding, region growing, and watershed provide the foundational groundwork for segmentation but require supplementary methods for effective noise reduction and boundary refinement.
5. Other Methods: The paper acknowledges unique approaches such as cellular automata and radial gradient index, which offer alternative segmentation strategies but are less prevalent due to computational overhead or reliance on ancillary input.

Implications and Future Directions

From a practical standpoint, the development of unconstrained BUS segmentation techniques is pivotal for medical imaging applications that vary significantly in quality and artifact presence. The paper advocates for creating a publicly accessible BUS image benchmark dataset, which can become a standard for evaluating and comparing segmentation technologies objectively.

Furthermore, the anticipated impact of deep learning techniques on BUS image segmentation is substantial. These methods—including CNNs and RNNs—could revolutionize the segmentation landscape by offering scalable and accurate models for automatic recognition and delineation of tumor regions.

Conclusion

This survey underscores the importance of advancing techniques that enhance the precision and reliability of BUS image segmentation. The future of breast ultrasound analysis lies in methodologies that can adapt to diverse imaging conditions while reducing resource footprints. Progress in this field is imperative for improving breast cancer diagnosis through real-time applications and ensuring accessibility in varied healthcare settings. The paper remains a valuable resource for experts aiming to expand and innovate within the scope of medical image segmentation.