Source-linked AI summary
Automatic Breast Ultrasound Image Segmentation: A Survey
Min Xian, Yingtao Zhang, H. D. Cheng, Fei Xu, Boyu Zhang, Jianrui Ding
TL;DR
Automatic BUS image segmentation must handle low SNR, artifacts, and large patient-to-patient breast-structure variation. This paper categorizes and reviews segmentation approaches, their principles, application issues, and advantages and disadvantages, concluding that GDMs have become more popular while robust performance across image settings remains a future direction.
Problem
Low SNR, artifacts, and large variations in breast structures make automatic BUS image segmentation challenging.
Method
The paper categorizes automatic BUS segmentation approaches and reviews their principles, application issues, advantages, and disadvantages.
Results
GDMs became more popular than traditional PDMs, while early techniques faded with the success of other approaches.
Takeaways & Limitations
Future work should pursue BUS segmentation techniques invariant to image settings.
Takeaways & Limitations
Current approaches work well under controlled conditions but their performance degrades greatly on other BUS images.
Abstract
from arXiv · showhide
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.
1. Introduction
Automatic BUS segmentation is essential for computer-aided diagnosis but remains difficult because ultrasound images are noisy, variable, and weakly constrained by anatomy. This survey organizes prior methods, discusses their trade-offs, and identifies future directions.
- Clinical role: BUS segmentation extracts tumors so quantitative features can support classification and diagnosis, making segmentation precision directly relevant to CAD performance.Features include tumor shape, size, and echo pattern.
- Challenges: BUS images are challenging because speckle noise, low contrast, low SNR, artifacts, and large inter-patient structural variation complicate segmentation.Tumor shape, size, and echo strength also vary substantially across patients, limiting the usefulness of fixed priors.
- Approach taxonomy: Existing approaches are classified as semi-automated or fully automated according to the degree of human intervention.Semi-automated methods may require an ROI, lesion seed, or initial boundary, whereas fully automated methods use breast-ultrasound and oncology priors.
- Approach taxonomy: The survey classifies BUS segmentation methods into six main categories, with graph-based, deformable-model, and other leading categories dominating the literature.The final three categories are described as classical image-processing approaches.
- Survey scope: Because single image-processing techniques are insufficient for this difficult task, most successful approaches combine hybrid techniques with biological priors.The survey reviews principles, advantages, disadvantages, application issues, and future directions.
2. Fundamental Issues of BUS Image Segmentation
BUS segmentation depends on handling speckle noise, preserving edges, modeling domain priors, and validating results appropriately. The survey also examines human interaction, automatic tumor detection, feature design, and learning-based strategies.
- Denoising and Preserving Edge: Speckle artifacts are a major BUS segmentation difficulty because they depend on tissue and cannot be effectively modeled.Denoising methods range from simple mean and Gaussian filters to speckle-specific diffusion and filtering approaches.
- Denoising and Preserving Edge: Mean and Gaussian filters are simple and fast but blur edges, whereas edge-based methods prefer denoising approaches that preserve edges to prevent segmentation leakage.The filter choice therefore depends on whether the segmentation approach is sensitive to image edges or gradients.
- Human Intervention: Semi-automatic methods can handle extremely difficult cases, but user interaction makes them operator-dependent, unreproducible, and costly for large datasets.Interaction intensity and sensitivity have not yet been evaluated in BUS segmentation work.
- Human Intervention: Fully automatic segmentation is operator-independent and reproducible, but tumor detection must first localize the lesion and initialize subsequent segmentation steps.Detection may output an ROI, rough boundary, seeds, or candidate regions.
- Priors and learning: Empirical domain-prior formulas can detect tumors effectively, yet predefined reference points limit robustness and flexibility.Suggested directions include adaptive reference points and learning-based formulations.
- Priors and learning: Learning-based approaches are promising and can improve by combining global and local features or learning deep breast-structure representations with convolutional neural networks.The survey also notes that ordinary segmentation frameworks alone cannot achieve good BUS performance.
- Biological and image priors: BUS segmentation models incorporate intensity, texture, edge, layer, topology, and smoothness priors to distinguish tumors from surrounding tissue.Texture and local-region features provide more descriptive power than intensity, while frequency-domain edge methods can reduce sensitivity to contrast.
3. Graph-based approaches
Graph-based BUS segmentation represents pixels or superpixels as connected graph nodes and combines image evidence with prior constraints through energy-based optimization. The survey covers MRF-MAP and graph-cut formulations, their parameterization and optimization choices, and practical limitations such as initialization sensitivity and implementation constraints.
- Graph representation: Graph-based methods organize pixels or superpixels as nodes connected by neighborhood-defined edges with nonnegative separation weights.The weights encode the cost of assigning different labels to connected nodes.
- MRF-MAP approaches: MRF-MAP segmentation combines likelihood and prior energies, then seeks the labeling that minimizes posterior energy.The likelihood models image observations, while the prior encodes configuration constraints.
- MRF-MAP approaches: Potts-model priors penalize label discontinuities between neighboring sites, while equal label priors make the unary prior term constant.The pairwise term is therefore commonly used to define the MRF prior energy.
- MRF-MAP approaches: BUS MRF-MAP methods commonly use ICM because it is faster than simulated annealing, but its greedy optimization is sensitive to initialization and local minima.Initialization may come from estimation or user interaction, and the issue is especially relevant for non-convex high-dimensional energies.
- Graph cuts: Graph cuts formulate segmentation as a minimum s-t cut whose cost combines data and smoothness terms, with max-flow providing the corresponding optimization.The data term can encode domain knowledge such as intensity, shape, and location, while smoothness penalizes neighboring discontinuities.
- Graph cuts: The survey identifies parameter learning and data-term and smoothness-term definition as central practical issues, while available implementations may require modification for BUS segmentation.Graph-cut formulations include terminal t-links and neighborhood n-links; one implementation cited by the survey is not directly for BUS segmentation.
4. Deformable models
Deformable models evolve curves or surfaces toward object boundaries using image-derived forces and prior knowledge. Parametric and geometric variants address different representation and topology needs, but BUS images require adaptations for weak boundaries and inhomogeneity.
- Overview: Deformable models move curves or surfaces toward object boundaries under forces defined from image information.They can incorporate edge, local-region, shape, and intensity priors.
- Model categories: Parametric models represent curves explicitly, whereas geometric models represent them implicitly as level sets of scalar functions.Geometric models can adapt to target topological changes.
- Parametric deformable models: PDM energy minimization balances internal continuity and smoothness terms with external image-feature costs that attract curves to boundaries.The weights α and β respectively penalize stretching and nonsmoothness, while P is image-feature based.
- Parametric deformable models: Default PDMs may fail to converge to concave boundaries or perform poorly when initialization is far from the minimum.Variants introduce different external energies to address these limitations.
- Geometric deformable models: GDMs use level-set evolution, speed functions, geometric measures, and image properties, requiring careful speed-function design and level-set initialization or reinitialization.Their topology adaptivity contributes to their greater popularity than traditional PDMs in BUS segmentation.
- BUS applications: Region-based deformable models can outperform edge-based models when BUS images contain substantial noise and weak object boundaries.Most successful approaches focus on improving GDMs for weak boundaries and image inhomogeneity.
5. Learning-based approaches
Learning-based BUS segmentation treats segmentation as pixel or superpixel classification and includes supervised and unsupervised methods. The survey reviews clustering, support-vector, neural-network, and Bayesian approaches, while highlighting the rise of deep learning.
- Overview: Learning-based BUS segmentation approaches classify pixels or superpixels using supervised or unsupervised learning.The section frames image segmentation as a classification problem.
- Unsupervised learning approaches: FCM assigns each data point membership values across clusters, enabling softer assignments and better handling of overlapped data than K-means.Its fuzzifier m controls cluster fuzziness; m = 1 yields a hard partition.
- Unsupervised learning approaches: FCM and K-means can converge only to local minima, and their results depend on initialization.Some BUS studies did not discuss how FCM membership values were initialized.
- Unsupervised learning approaches: Neutrosophic l-means incorporates membership indeterminacy and is reported to handle uncertainty in BUS images better.This extension targets blur boundaries and uncertainty in ultrasound images.
- Supervised learning approaches: Supervised methods include SVMs, ANNs, and naive Bayesian classifiers, using image, texture, morphology, intensity, and spatial features.Reported pipelines combine classifiers with candidate-region generation or region refinement.
- Summary: Learning-based approaches thrived during the last decade, and the survey identifies deep CNN and recurrent neural-network techniques as emerging directions.Many BUS approaches achieve good performance with supervised learning.
6. Classical Approaches: thresholding, region growing, and watershed
Thresholding, region growing, and watershed are simple classical approaches used for initial BUS segmentation. Their speed and simplicity often require preprocessing and more delicate refinement to achieve good performance.
- Thresholding: Thresholding groups pixels into regions using one or more thresholds based on intensity or other pixel features.Global thresholds are constant across an image, whereas adaptive thresholds vary with local features.
- Thresholding: Global thresholding is fast and effective when object and background intensity distributions are distinct, but degrades with low contrast, noise, or varying illumination.It is often used as a BUS pre-segmentation step.
- Region growing: Region growing expands seed regions using predefined intensity or local-feature criteria until no additional pixel satisfies the condition.Seeds may be selected interactively or generated automatically using region and pixel statistics.
- Watershed: Watershed can produce more stable results than thresholding and depends critically on marker selection.Local-minimum-gradient markers can cause noise-driven over-segmentation, requiring region merging.
- Overview: The section reviews thresholding, region growing, and watershed as classical BUS segmentation approaches.These methods are commonly combined with other methods.
- Summary: Classical approaches are simple, fast, and efficient for initial segmentation, but preprocessing and refinement are usually needed for good BUS performance.Preprocessing denoises and enhances contrast, while later methods refine the segmentation.
7. Other approaches
Other BUS segmentation approaches include cellular automata, cell competition, and radial gradient index methods. They offer specialized capabilities but face computational cost, interaction, or image-quality constraints.
- Cellular automata: Cellular automata represent image cells with states, neighborhoods, and transition rules that update cell states from neighboring states.BUS variants can integrate global transition-chain information and local texture correlation.
- Cellular automata: CA-based approaches support multiple-object segmentation and can generate precise boundaries without the shrink problem.They may also support user input online.
- Cellular automata: CA segmentation can have high computation cost, especially for large images.A fast CA-based approach is cited as a response to this limitation.
- Cell competition: Cell competition generates watershed-based cells, then lets adjacent cells split or merge into new regions through two competition types.The approach has been extended to 3D BUS segmentation with graph cuts for final tumor boundaries.
- Cell competition: Cell competition is simple and fast but requires substantial user interaction to select regions of interest or tumor regions.Its competition mechanism does not integrate task-related knowledge.
- Radial gradient index: RGI is easy to implement but computationally expensive and unable to obtain accurate BUS tumor boundaries when image quality is low.Its dependence on image gradients limits performance in such images.
8. Conclusions and Future Directions
The survey reviews automatic BUS segmentation approaches and identifies robustness, public resources, deep learning, and efficiency as key future directions. It emphasizes that current methods often work in controlled settings but degrade across variable image conditions, while practical deployment also requires lower resource demands.
- Conclusions: The survey reviews automatic BUS image segmentation approaches and compares four main categories.A general comparison of the four main categories is listed in Table 6.
- Unconstrained BUS image segmentation techniques: Most existing methods perform well on controlled BUS images but degrade greatly when image quality, artifacts, or tumor counts vary.Controlled settings include high contrast, fewer artifacts, and one tumor per image.
- Unconstrained BUS image segmentation techniques: Unconstrained segmentation techniques invariant to image settings are considered crucial for applying BUS computer-aided diagnosis systems in clinical practice.The survey identifies learning invariant and discriminative tumor representations as a potential direction.
- Benchmark: Publicly accessible BUS benchmarks could support fairer comparisons, novel strategies, better approaches, and advances in breast cancer research.Constructing such a dataset would require many years of work and substantial resources, but its impact could be significant and profound.
- Deep learning: Deep learning has potential for BUS segmentation because it can learn compact image representations from sufficiently large datasets.The survey relates this direction to deep learning’s performance across recognition, classification, semantic segmentation, and medical applications.
- High performance segmentation: Reducing computation and memory demands is necessary for portable, real-time BUS devices and affordable early breast-cancer detection.Many current approaches are computation- and memory-intensive, limiting widespread application, especially in resource-limited regions.