Portal de Periódicos da CAPES

Adaptive localised region and edge‐based active contour model using shape constraint and sub‐global information for uterine fibroid segmentation in ultrasound‐guided HIFU therapy

2017; Institution of Engineering and Technology; Volume: 11; Issue: 12 Linguagem: Inglês

10.1049/iet-ipr.2016.0651

1751-9667

Xiangyun Liao, Zhiyong Yuan, Qianqian Tong, Jianhui Zhao, Qiong Wang,

AI in cancer detection

IET Image ProcessingVolume 11, Issue 12 p. 1142-1151 Research ArticleFree Access Adaptive localised region and edge-based active contour model using shape constraint and sub-global information for uterine fibroid segmentation in ultrasound-guided HIFU therapy Xiangyun Liao, Xiangyun Liao Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, 518055 Guangdong, People's Republic of China School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorZhiyong Yuan, Corresponding Author Zhiyong Yuan zhiyongyuan@whu.edu.cn School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorQianqian Tong, Qianqian Tong School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorJianhui Zhao, Jianhui Zhao School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorQiong Wang, Corresponding Author Qiong Wang wangqiong@siat.ac.cn Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, 518055 Guangdong, People's Republic of ChinaSearch for more papers by this author Xiangyun Liao, Xiangyun Liao Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, 518055 Guangdong, People's Republic of China School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorZhiyong Yuan, Corresponding Author Zhiyong Yuan zhiyongyuan@whu.edu.cn School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorQianqian Tong, Qianqian Tong School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorJianhui Zhao, Jianhui Zhao School of Computer, Wuhan University, Wuhan, 430072 Hubei, People's Republic of ChinaSearch for more papers by this authorQiong Wang, Corresponding Author Qiong Wang wangqiong@siat.ac.cn Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, 518055 Guangdong, People's Republic of ChinaSearch for more papers by this author First published: 12 October 2017 https://doi.org/10.1049/iet-ipr.2016.0651Citations: 10AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract Uterine fibroids segmentation in ultrasound images is of great importance in the definition of intra-operative planning of ultrasound-guided high-intensity focused ultrasound (HIFU) therapy. However, it is challenging to obtain accurate, robust and efficient uterine fibroid segmentation due to low quality of ultrasound images. In this study, the authors propose a novel adaptive localised region and edge-based active contour model using shape constraint and sub-global information to accurately and efficiently segment the uterine fibroids in ultrasound images with robustness against initial contour. The authors first define adaptive local radius for the localised region-based model and combine it with the edge-based model to accurately and efficiently capture image's heterogeneous features and edge features. Then, they incorporate a shape constraint to reduce boundary leakage or excessive contraction to obtain more accurate segmentation. To overcome the initialisation sensitivity, they introduce the sub-global information to prevent the curve from trapping into the local minima and obtain robust results. Furthermore, the authors optimise computation by adaptively sharing local region and employing the multi-scale segmentation method to achieve efficient segmentation. The proposed method is validated by uterine fibroid ultrasound images in HIFU therapy and the results demonstrate that it can achieve accurate, robust and efficient segmentation. 1 Introduction Uterine fibroids are common benign myometrial neoplasms which can cause heavy or prolonged menstrual bleeding and resultant anaemia in women of reproductive age [1, 2]. Ultrasound-guided high-intensity focused ultrasound (HIFU) therapy, as shown in Fig. 1, is a non-invasive surgery which can effectively cure the uterine fibroids by focusing external high-intensity energy on therapeutic target to ablate the uterine fibroids [3, 4]. The accurate, robust and efficient uterine fibroids segmentation in the guidance ultrasound images of HIFU therapy is essential for the definition of intra-operative planning of the surgery [5]. Unfortunately, the ultrasound image in HIFU therapy is of poor quality with severely low signal-to-noise ratio (SNR), weak boundary and heterogeneous image intensities due to the interference of deaerated water when imaging by type-B ultrasound during the treatment, which brings enormous difficulties in accurately and efficiently segmenting uterine fibroids with robustness [6-8]. Fig. 1Open in figure viewerPowerPoint Ultrasound-guided HIFU therapy (a) Patient in the prone position on the ultrasound-guided HIFU equipment, (b) Uterine fibroid ultrasound image in HIFU therapy, it is of poor quality with low SNR, weak boundary and heterogeneous intensities For the past decades, the problem of image segmentation has been extensively exploited in the literature and variety of methods has been developed to provide satisfactory segmentation results [9-11]. Among these methods, active contour model is a popular and widely used energy-based segmentation method, which allows a contour to deform so as to minimise a given energy functional in order to produce the desired segmentation contour with smooth and closed features [12-15]. Active contour model was firstly proposed by Kass et al. [16], and it can be classified into two main categories based on the information used for segmentation: edge-based model [17] and region-based model [18]. The edge-based model uses image gradient information to drive the active contour move towards the target region boundaries, such as the geodesic active contours (GAC) and the gradient vector flow, which can easily capture image edge information and have the advantage of fast convergence [19, 20]. However, the gradient information is highly localised image information which will lead to sensitivity to image noise and initial contour, and will easily produce obvious boundary leakage when applied to ultrasound images with weak boundary. Region-based model adopts statistics information of forehead and background regions to drive the contour move towards the boundaries by minimising the given functional, and it can obtain better segmentation performance for images with weak boundaries than edge-based model [21]. The C–V model, which is the most well-known and widely used region-based active contour model, uses global region statistics information with the assumption that image intensities are homogeneously distributed [22]. However, the C–V model is not ideal for segmenting images with heterogeneous intensities by global statistics information and can easily lead to erroneous segmentation results. To tackle this problem, many methods were developed by using local statistics information for achieving good segmentation in images with heterogeneous intensities [23, 24]. Lankton and Tannenbaum [25] proposed the localised CV (LCV) model which adopted local rather than global image statistics and evolved a contour based on the local region information. The LCV model is capable of segmenting objects with heterogeneous intensities which is difficult to capture correctly using a standard global method. However, the LCV model suffers from obvious boundary leakage or excessive contraction for images with severely weak boundaries and low efficiency due to the separately local energy minimisation for every point in the curve. Li et al. [26] introduced a local binary fitting (LBF) energy with a kernel function, which enabled the extraction of accurate local image information and can segment images with heterogeneous intensities. The above two models use local statistics information and have better segmentation capacity for images with heterogeneous intensities than global region-based models. Unfortunately, for uterine fibroid ultrasound images in HIFU therapy, they still produce erroneous segmentations with boundary leakage or excessive contraction due to severely poor quality of the images. Moreover, the segmentations by the above two methods are inefficient as they need to separately minimise the energy in the local region for each point on the evolution curve. Besides, some researchers combined the region-based model with the edge-based model to obtain better performance in image segmentation. Zhang et al. [27] proposed the selective binary and Gaussian filtering regularised level set (SBGFRLS) method by fusing the GAC model and the C–V model to accurately and effectively segment the homogeneous images. While the SBGFRLS model uses global statistics information which leads to difficulties in locating the exact object boundaries, especially for the ultrasound images with heterogeneous intensities. Tian et al. [28] proposed a novel active contour model by integrating the GAC model, the LBF model and the SBGFRLS model to segment objects in magnetic resonance images. However, it produces incorrect segmentations when segmenting ultrasound images with weak boundary and low contrast. Recently, there are many shape-based constraints considered in the active contour models. Wu et al. [29] presented a fast external force named gradient vector convolution for the parametric active contour model and incorporated the circle-based energy into it to segment the left ventricle myocardial boundary. Tran et al. [30] presented a fuzzy energy-based active contour model with shape prior for image segmentation, where the shape term is defined as the distance between the evolving shape and a reference one, constrains the evolving contour with respect to the reference shape. A multi-scale shape constrained LCV (MSLCV) model was proposed to obtain good segmentation in ultrasound images and it can reduce boundary leakage or excessive contraction by the shape constraint [31]. However, the performance of these models in ultrasound images is sensitive to the wake boundary as well as the initialisation of the constrained shape, and the segmentation is still inefficient for the uterine fibroid segmentation in ultrasound-guided HIFU therapy. In the present work, we propose a novel adaptive localised region and edge-based active contour model using shape constraint and sub-global information for obtaining accurate, robust and efficient segmentation of uterine fibroid ultrasound images in HIFU therapy. Our method relies on the combination of the adaptively localised region, edge information, shape constraint and sub-global information to achieve accurate and robust segmentation results. In addition, our method can efficiently segment the ultrasound images by adopting the multi-scale scheme and optimising the computation in localised region. Our contributions are summarised as follows: To achieve accurate segmentation results, we propose an adaptively localised region and edge-based model with shape constraint by combining the merits of the LCV model and the GAC model, adaptively changing the local radius to accurately and efficiently capture the images' heterogeneous features and edge features and incorporating a shape constraint to reduce boundary leakage and excessive contraction during the curve evolution. To overcome the initialisation sensitivity and improve the segmentation efficiency, we introduce the sub-global statistics information to prevent the evolved contour from trapping into the local minima and thus achieve robust segmentations. We optimise the computation by adaptively sharing localised region for neighbouring points and employing multi-scale segmentation method to achieve efficient segmentations. The remainder of this paper is organised as follows. Section 2 elaborates the proposed method. Section 3 gives the experimental results and Section 4 draws conclusion. 2 Method 2.1 Overview Based on the combination of the GAC model [19] and the LCV model [25] with adaptively changed local radius, we propose to incorporate a shape constraint, introduce the sub-global information and optimise the computation to finally form the proposed method which is accurate, robust and efficient in segmenting uterine fibroid ultrasound images in HIFU therapy. The schematic illustration of the proposed method is illustrated in Fig. 2. Fig. 2Open in figure viewerPowerPoint Schematic illustration of the proposed method 2.2 Adaptive localised region and edge-based model Accurately segmenting uterine fibroid ultrasound images in HIFU therapy is a challenging task due to the images' severely low SNR, weak boundary and heterogeneous intensities. It is essential to effectively extract the useful features from the images to help us segment the uterine fibroids. Considering the advantages that the LCV model is capable of segmenting heterogeneous objects and the GAC model can easily capture the edge information, we propose a localised region and edge-based model which combines the merits of the LCV model and the GAC model to effectively capture the heterogeneous features and edge features for accurately and efficiently segmenting the uterine fibroid ultrasound images. The localised region and edge-based model simultaneously makes use of both heterogeneous features and edge features, which helps the curve evolve to the target boundaries. For a given image I defined on the domain , the closed contour C can be represented as zero level set of a signed distance function , i.e. . The evolution equation of the level set function is as follows: (1)where is the smoothed version of Dirac function [25], and are the terms related to the LCV model [25] and the GAC model [19], and they are defined as (2) (3)where is the characteristic function to define a local region for each point on the evolution curve, and are, respectively, the mean intensities in the interior and exterior of the contour localised by at a point x [25]. The arclength of the curve is penalised with weight to keep the curve smooth, is a constant, and g is the edge stopping function [19]. It is worth noting that the local radius greatly impacts the accuracy and efficiency of the segmentation as the curve uses the information in local regions to evolve and the computational complexity depends on the size of the local region. The segmentation becomes inaccurate and inefficient when the local radius is too large, which brings redundant information and much computation in the local region. While the segmentation becomes inaccurate when the local radius is too small, which cannot provide sufficient information to obtain accurate segmentation. Thus, to obtain more accurate and efficient segmentation, we propose the adaptive localised region and edge-based model which adaptively selects the local radius for each point in the curve. The adaptive local radius for point x in the curve is selected according to the dynamic evaluation of the region near point x during curve evolution: (4)where is the initial local radius which is defined in initialisation [(10) in Section 2.3], and are, respectively, the mean intensities in the interior and exterior of the contour localised by at a point x. Fig. 3 illustrates the adaptive local radius during curve evolution, different points in the curve have different local radius. Fig. 3Open in figure viewerPowerPoint Adaptive local radius during curve evolution (a) Adaptive local radius, (b) Different local radius for each point in the curve. Green circle represents the local region with radius , and are mean intensities for the local regions inside and outside the contour, respectively By adaptively changing the local region radius, we propose the adaptive localised region and edge-based model in order to better capture heterogeneous features and edge features for improving segmentation performance. Unfortunately, region and edge features are still not sufficient to obtain reliable and precise segmentation for ultrasound images [6]. Specifically, it is even more challenging for uterine fibroid ultrasound images in HIFU therapy as these images have lower quality than ordinary ultrasound images due to the interference of deaerated water when imaging by type-B ultrasound. For the regions with severely weak boundaries in uterine fibroid ultrasound images in HIFU therapy, the adaptive localised region and edge-based model still produces erroneous segmentation with boundary leakage or excessive contraction which is a problematic issue that needs to be resolved. 2.3 Shape constraint In order to reduce boundary leakage and excessive contraction, we incorporate a shape constraint into the adaptive localised region and edge-based model to improve the segmentation accuracy in regions with weak boundaries. Considering the fact that the shape of uterine fibroids approximates an ellipsoid, we use an ellipse without training as the initial contour as well as the shape constraint which impacts the curve evolution by adding shape constraint energy to each point's energy. The shape constraint energy is acquired by a function of the nearest distance between the point and the initial contour (5)where is a positive constant and is the shape constraint force, represented as (6)where is the position of point x in the curve, is the position of the nearest point to point x in the curve, is the level set representation of the shape constraint, is the sign function which determines the force direction which is towards the shape constraint contour to reduce boundary leakage or excessive contraction. The level set evolution equation for shape constraint is as follows: (7) (8)After adding the shape constraint to the adaptive localised region and edge-based model, the corresponding level set evolution equation is (9)Suppose the maximal and minimal coordinates of the initial contour in the x-axis and y-axis are and , respectively, here we set the initial local radius as (10)The shape constraint is illustrated in Fig. 4a. We use the shape constraint to reduce boundary leakage and excessive contraction in the regions with weak boundaries to achieve better segmentation accuracy. However, the shape constraint has a natural shortcoming of initialisation sensitivity. The curve evolution is significantly affected by the quality of the initial contour that the shape constraint causes the curve to evolve slowly and trap into the local minima in local regions with weak boundaries. Fig. 4Open in figure viewerPowerPoint Shape constraint and the sub-global region (a) Shape constraint. The yellow ellipse is the initial contour, the red curve is the evolution curve, the green arrow represents the shape constraint forces to reduce boundary leakage, and the blue arrows represent the shape constraint forces to reduce excessive contraction, (b) Sub-global region. The region inside the red rectangle is the sub-global region 2.4 Sub-global information To overcome the initialisation sensitivity and obtain robust segmentation, we propose to introduce the sub-global information to promote the curve evolution continuously towards the target boundaries in local regions with low contrast and prevent local minima of energy. Fig. 4b illustrates the sub-local region which is slightly larger than the initial contours, and is defined as follows: (11)where and are the maximal and minimal coordinates of the initial contour in the x-axis and the y-axis, respectively, and , . We define the sub-global region energy as follows: (12)In (12), is the sub-global force function, which is defined as (13)where are, respectively, defined as (14)Taking the first variation of with respect to , we obtain the following evolution equations: (15) (16)Here we define a characteristic function to determine where to apply the sub-global information according to the absolute ratio of the mean intensities in the interior and exterior of the contour, respectively. When is very close to , we treat this local region as the low contrast region and apply the sub-global information to promote the curve evolution: (17)where is a constant and usually we set . By using sub-global information, we finally obtain the corresponding level set evolution equation: (18) 2.5 Computation optimisation We have combined the LCV model and GAC model with adaptive local radius, incorporated a shape constraint and introduced sub-global information for achieving accurate and robust segmentation results. Unfortunately, the curve evolution is still inefficient due to iterative computation of separate energy minimisation in local region, shape constraint and sub-global information processing. This drawback limits application of our method in segmenting uterine fibroids in ultrasound images during the treatment. The main computation in segmentation comes from the energy minimisation in local region for each point in the curve and the iterative processing of the whole pipeline. In these regards, we optimise the computation by two ways and finally form the proposed method. One way is to adaptively share the same local regions for a certain continuous points in the curve, thus it is unnecessary to define a local region and calculate its minimal energy for every point in the curve. Starting from point x in the evolution curve, we define as the number of continuous points in the curve which are to share the same local region: (19)where is the adaptive local radius as shown in (4), represents the maximum integer that is less than x, is a constant and usually we set . The other way to improve segmentation efficiency is to employ the multi-scale segmentation method [31]. The basic idea is that we adopt the Gaussian pyramid algorithm to decompose the original ultrasound image into different scale images and from fine to coarse, and perform coarse segmentation on the coarse-scale image by the adaptive localised region and edge-based model using shape constraint and sub-global information. Then we adopt the segmentation result as an initial contour for the fine-scale image, thus gradually optimising the contour and reaching the final segmentation result efficiently. Fig. 5 illustrates the local regions sharing and the multi-scale segmentation method. Fig. 5Open in figure viewerPowerPoint Local region sharing and multi-scale segmentation method (a) Continuous points sharing the local region, (b) Multi-scale segmentation, the resolution of image and are , and , respectively 3 Experiments We conducted a series of experiments on uterine fibroid ultrasound images in HIFU therapy to illustrate the effectiveness of the proposed method. We validated the proposed method step by step and compared it with the GAC [19], the LCV [25], the SBGFRLS [27], Tian's method [28], Wu's method [29], Tran's method [30] and the MSLCV [31]. The uterine fibroid images used in the experiments came from the HIFU centre of the second affiliated hospital of Chongqing Medical University. For all experiments in this section, the green contours are manual segmentation results by a specialist as ground truth, while the red contours are computerised segmentation results. The image resolution is . The iteration of the curve evolution is 500. For multi-scale segmentation, we set iterations for images , and are, respectively, 100, 100 and 300. All experiments were conducted on a PC equipped with Intel(R) Xeon(R) E5-2640 v3 CPU (2.60 GHz), largest available memory of 64 GB RAM and MATLAB 2012. In this paper, we adopt dice similarity coefficient (DSC) [32] and mean sum of square distance (MSSD) [32] to quantitatively evaluate the segmentation performance. The DSC and MSSD are, respectively, defined as (20) (21)where S is the computerised segmentation result, R is the reference (ground truth), represents the overlap of S and R. is the area operator and represents the area of , and N is the number of points in contour S. The DSC value measures the similarity of two contours, and the closer the DSC value is to 1, the better the computerised segmentation fits to the ground truth. The MSSD measures the average minimum Euclidean distance square between two contours, and the smaller the MSSD value is, the more accurate the computerised segmentation is. 3.1 Evaluation of the proposed method The formulation of the proposed method is actually made up of four components, there are in turn: (i) adaptive localised region and edge-based model, (ii) shape constraint, (iii) sub-global information and (iv) computation optimisation. In the following experiments, we evaluate the above four components step by step. 3.1.1 Evaluation of the adaptive localised region and edge-based model Fig. 6 demonstrates the segmentation results of a uterine fibroid ultrasound image in HIFU therapy using four methods: GAC, LCV, localised region and edge-based model (without adaptive local radius), adaptive localised region and edge-based model. The performance of each method is shown in Table 1. Table 1. Segmentation performance in Fig. 6 with 500 iterations Methods DSC MSSD Time, s 1. GAC 0.826 433.02 15.05 2. LCV 0.898 147.56 50.82 3. Localised region and edge-based model 0.918 80.22 61.43 4. Adaptive localised region and edge-based model 0.938 41.26 57.48 Fig. 6Open in figure viewerPowerPoint Segmentation results using different methods with 500 iterations. Red contours are computerised segmentation results, green contours are ground truth (a) Segmentation result of the GAC model, , (b) Segmentation result of the LCV model, and initial local radius , (c) Segmentation result of the localised region and edge-based model, , , initial local radius , (d) Segmentation result of the adaptive localised region and edge-based model, , and initial local radius It can be observed from Fig. 6a that the GAC model can evolve to the boundaries in regions with relatively high contrast, while it produces obvious boundary leakage in low contrast regions. As shown in Fig. 6b and Table 1, the LCV model (DSC = 0.898, MSSD = 147.56) can better process the heterogeneous intensity and achieves more accurate segmentation than the GAC model (DSC = 0.826, MSSD = 433.02), but it leads to boundary leakage and excessive contraction in regions with weak boundaries, and is inefficient (time = 50.82 s) due to the separate local energy minimisation for every point in the curve. As shown in Fig. 6c and Table 1, the localised region and edge-based model without adaptive local radius (DSC = 0.918, MSSD = 80.22) can better capture the heterogeneous features and edge features and achieves more accurate segmentation results than the GAC model and the LCV model. However, in regions with low contrast, it produces obvious excessive contraction. Fig. 6d and Table 1 demonstrate the segmentation results of the adaptive localised region and edge-based model (DSC = 0.938, MSSD = 41.26), which achieves more accurate and more efficient (time = 57.48 s) segmentation results than the localised region and edge-based model (time = 61.43 s) due to the adaptive local radius which constructs proper local regions for each point. 3.1.2 Evaluation of shape constraint Although the adaptive localised region and edge-based model improves the accuracy of the segmentation, as shown in Fig. 6d, it causes obvious boundary leakage and excessive contraction in regions with weak boundary. To resolve this problem, we use an ellipse as the shape constraint and incorporate it into the adaptive localised region and edge-based model by adding shape constraint energy to each point's energy. Fig. 7 and Table 2 illustrate the segmentation results by using the adaptive localised region and edge-based model without or with shape constraint. Fig. 7a shows a good initial contour placed roughly on the centre of the uterine fibroid. It can be observed from Fig. 7 and Table 2 that the adaptive localised region and edge-based model using shape constraint (DSC = 0.952, MSSD = 30.40) achieves more accurate segmentation result with less boundary leakage than the adaptive localised region and edge-based model without shape constraint (DSC = 0.938, MSSD = 41.26). The experimental result demonstrates the effectiveness of the shape constraint. Table 2. Segmentation performance in Fig. 7 with 500 iterations Methods DSC MSSD Time, s 1. Adaptive localised region and edge-based model without shape constraint 0.938 41.26 57.48 2. Adaptive localised region and edge-based model with shape constraint 0.952 30.40 69.74 Fig. 7Open in figure viewerPowerPoint S