Original researchFully automatic Breast ultrasound image segmentation based on fuzzy cellular automata framework
Introduction
Ultrasound (US) image segmentation plays an important role in computer-aided diagnosis (CAD) systems [1]. The goal of US image segmentation is to select the target areas to provide the necessary information for the step of feature extraction or lesion classification [2], [3], [4]. Different model-based segmentation approaches have been proposed and shown advantages in dealing with different kinds of ultrasound images [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]. However, in some kinds of ultrasound images, the boundaries between foreground regions and background regions are quite blurry and the structures of the target areas are quite complex, which may result in the actual target regions (lesion regions) being wrongly partitioned into normal tissue regions. Existing segmentation methods usually formulated the energy functions according to the intensity distribution of different regions and optimized to the local minimum. Therefore, in most of existing segmentation approaches, the energy function and initial condition need to be well formulated. If the initial segmentation condition is formulated unsuitable (e.g. is easily affected by noise), the process may lead to an unexpected result. However, in some cases, the intensity distributions between different lesion regions are quite similar, leading some parts of boundaries quite blurry. Without spatial information, different areas are evaluated equally, and pixels of blurry boundaries may not be processed effectively. In addition, due to speckle noise, the signal/noise ratio of ultrasound image is quite low, and the structures and details of the tissues are quite fuzzy, which may affect segmentation result seriously, i.e., over-segmentation may happen. Existing approaches usually utilize noise reduction methods for suppressing the noise [9]. However, the details of US images may be damaged after noise-reduction operation [15]. Segmentation of breast ultrasound (BUS) images is still a difficult task Cellular automata (CA) can achieve highly complicated models with relative simple initial conditon and evolution laws [16]. The basic concept of cellular automata is to evaluate the variances of elements to evolute according to a pre-programed rule operating on a given neighborhood system [17], [18]. For improving the robustness of segmentation to the initial condition, [19] proposed an interactive image segmentation approach based on CA. By comparing the features of neighboring pixels (e.g. intensity or Euclidean distance), this approach characterized each pixel’s class according to the relationship of their feature distances and shown the advantages in handling the images with uniform intensity distribution. However, to the images with non-uniform intensity distribution, this approach may be affected by irregular or rapid changes of intensities caused by speckle noise, non-homogeneous echoes in US image, leading the energy transition of CA becoming unsteady and resulting the segmentation converge in an unsuitable position. For overcoming this problem, we introduced a segmentation approach based on cellular automata for US image segmentation in our previous study [2]. By establishing the local area similarity metrics by comparing the difference of intensities between neighboring pixels and the differences of Haralick textures [20] between different local areas, this approach produced a good result in dealing with the high speckle noise. However, the similarity metrics used in its energy function were sensitive to the significant changes of the intensities, which may cause the target boundaries wrongly generated. Since the goal of segmentation is to extract the global contour of target regions, the details in different target regions should be leaved out even their intensities have large difference. In addition, existing CA-based segmentation approaches require the seed pixel to be manually selected in advance, and fully segmentation of ultrasound images have not been extensively evaluated or validated [8], [16]. In approach [2], the seed pixels were generated according to the size of region of interest produced by the users. However, to the cases with complicate structures, the seed pixels were difficult to cover the crucial regions with irregular branch shapes. Moreover, traditional CA based segmentation approaches usually represented the cells’ states with certain values. In real US image segmentation applications, there were many uncertain ascription problems of pixels (e.g. blurry boundaries), which, for the reasons above, may not be desirable in practical US image processing applications.
In this paper, a fully automatic segmentation approach based on fuzzy cellular automata (FCA) framework is proposed for medical image segmentation. Different from existing CA based segmentation approaches, in proposed method, not only evolution of FCA is automatic, but the seed points are also selected automatically by developing an automatic seed point templates generation strategy that considers both intensity and non intensity distribution information comprehensively. In the evolution process, the certain outputs of element classes in traditional CA are transformed into the fuzzy decision expression, and a majority voting strategy is employed for defuzzification. To handle the non-homogeneous intensity variation caused by the high speckle noise or non-homogeneous echoes in US images, two kinds of texture features are extracted by considering both intensity distribution and non intensity distribution of pixels in different regions. In the energy transition function of proposed method, both spatial information and non spatial information of image are considered simultaneously and integrated into the energy comparison function. The spatial information is calculated according to the pixel’s median in a given neighborhood. For modeling the non spatial image information, both intensity distribution and structural similarity relation between pixels and areas are considered. In the experiment, a series of ultrasound images with speckle noise, complicate structures and blurry boundaries are utilized for validating the performance of the proposed approach, and the results segmented by our method. Furthermore, other segmentation approaches are also compared to the corresponding target regions marked by the radiologist. The segmentation results show that the proposed approach can effectively process the US images with blurry boundaries and complex irregular shapes, and is also robust to the noise.
The structure of the paper is organized as follows. In Section 2, the proposed method is introduced. The segmentation results of proposed method are carried out and discussed in Section 3. Finally, a brief conclusion is given in Section 4.
Section snippets
Data acquisition
In our experiment, a series of breast US (BUS) images (125 benign and 110 malignant) are utilized for validating the segmentation performance of the proposed approach, while considering the characteristics of irregular and complicated shape, low signal/noise ratio, low contrast and blurry boundaries in BUS images. The BUS images were obtained from the Second Affiliated Hospital of Harbin Medical University (Harbin, China) collectedfrom HITACHI Vision 900 system (Hitachi Medical System, Tokyo,
Results
To test the segmentation performance, the proposed approach is applied to a series of BUS images, and the LSAC method and the MRF-based method are applied to the same images for comparison. Fig. 2(a) is an original BUS image, where it is seen that some normal tissues have the similar intensity distribution with the tumor, leading some parts of boundaries between lesion and normal tissue regions being quite blurry. The segmentation results the LSAC method, the MRF-based method and the proposed
Conclusion
In this paper, a fully automatic segmentation approach based on fuzzy cellular automata is proposed for handling the difficulty in dealing with the uncertain problems during the segmentation process of breast ultrasound images. For the initial condition of segmentation, the seed cells are selected automatically by exploring an automatic seed point templates generation method by integrating different image information more comprehensively. For handling the fuzzy information and uncertainty
Acknowledgements
This work is supported by the China Postdoctoral Science Foundation funded project (Grant NO. 2015M581450), Program for Innovation Research of Harbin City (Grant NO. 2015RAQXJ090), Education Project of Heilongjiang Provincial Department (Grant NO. 12541420), State Ley Laboratory of Intelligent Manufacturing System Technology.
References (26)
- et al.
Computer-aided diagnosis system for Breast cancer using B-Mode and color Doppler flow images
Opt. Eng.
(2012) - et al.
An early vision-based snake model for ultrasound image segmentation
Ultrasou. Med. Biol.
(2000) - et al.
Probability density difference-based active contour for ultrasound image segmentation
Pattern Recogn.
(2010) - et al.
Fully automatic segmentation of Breast ultrasound images based on Breast characteristics in space and frequency domains
Pattern Recogn.
(2015) - et al.
Automatic identification of the optimal reference frame for segmentation and quantification of focal liver lesions in contrast-Enhanced ultrasound
Ultrasou. Med. Biol.
(2017) - et al.
Completely automated segmentation approach for breast ultrasound images using multiple-domain features
Ultrasound Med. Biol.
(2012) - et al.
Automated segmentation of ultrasonic Breast lesions using statistical texture classification and active contour based on probability distance
J. Digit. Imaging
(2009) - et al.
Automatic ultrasound segmentation and morphology based diagnosis of solid breast targets
Breast Cancer Res. Treat.
(2005) - et al.
An effective approach of Lesion segmentation within the breast ultrasound image based on cellular automata
J. Digit. Imaging
(2012) - et al.
Computer aided diagnosis system for Breast cancer based on color doppler flow imaging
J. Med. Sys.
(2012)
Breast ultrasound image segmentation: a survey
Int. J. Compu. Assis. Radiol. Surgery
Ultrasound image segmentation and tissue characterization Proceedings of the Institution of Mechanical Engineers
Part H: J. Eng. Med.
A Systematic Performance Evaluation of Interactive Image Segmentation Methods Based on Simulated User Interaction. In 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro
Cited by (15)
Breast Cancer Images Segmentation using Fuzzy Cellular Automaton
2023, Procedia Computer ScienceDeep Cellular Automata-Based Feature Extraction for Classification of the Breast Cancer Image
2023, Applied Sciences (Switzerland)Trustworthy Breast Ultrasound Image Semantic Segmentation Based on Fuzzy Uncertainty Reduction
2022, Healthcare (Switzerland)WEIGHTED AREA CONSTRAINTS-BASED BREAST LESION SEGMENTATION IN ULTRASOUND IMAGE ANALYSIS
2022, Inverse Problems and ImagingBreast Ultrasound Image Segmentation Algorithm Using Adaptive Region Growing and Variation Level Sets
2022, Mathematical Problems in EngineeringA Complex Background Image Registration Method Based on the Optical Flow Field Algorithm
2022, Communications in Computer and Information Science