A dynamic graph-cuts method with integrated multiple feature maps for segmenting kidneys in 2D ultrasound images

Abstract

Rationale and Objectives

Automatic segmentation of kidneys in ultrasound (US) images remains a challenging task due to high speckle noise, low contrast, and large appearance variations of kidneys in US images. Since texture features may improve the US image segmentation performance, we propose a novel graph-cuts method to segment kidney from US images by integrating image intensity information and texture feature maps.

Materials and Methods

We develop a new graph cuts based method to segment kidney US images by integrating original image intensity information and texture feature maps extracted using Gabor filters. To handle large appearance variation within kidney images and improve computational efficiency, we build a graph of image pixels close to kidney boundary instead of building a graph of the whole image. To make the kidney segmentation robust to weak boundaries, we adopt localized regional information to measure similarity between image pixels for computing edge weights to build the graph of image pixels. The localized graph is dynamically updated and the graph cuts based segmentation iteratively progresses until convergence. Our method has been evaluated based on kidney US images of 85 subjects. The imaging data of 20 randomly selected subjects were used as training data to tune parameters of the image segmentation method, and the remaining data were used as testing data for validation.

Results

Experiment results demonstrated that the proposed method obtained promising segmentation results for bilateral kidneys (average Dice index=0.9446, average Mean Distance=2.2551, average Specificity=0.9971, average Accuracy=0.9919), better than other methods under comparison (p<0.05, paired Wilcoxon rank sum tests).

Conclusion

The proposed method achieved promising performance for segmenting kidneys in 2D US images, better than segmentation methods built on any single channel of image information. This method will facilitate extraction of kidney characteristics that may predict important clinical outcomes such progression chronic kidney disease.

INTRODUCTION

Ultrasound (US) imaging has been widely used to examine kidneys for structural abnormalities and to measure kidney anatomical features such as renal parenchymal area that have been associated with development of end stage renal disease (1–3). However, automatic segmentation of kidneys in 2D US images remains a challenging task due to high speckle noise and low contrast between foreground and background, as well as weak boundaries and large appearance variations of kidneys in 2D US images (4). Automatic segmentation of ultrasound images of kidneys will facilitate extraction and quantification of anatomical features such as renal parenchymal area and kidney echogenicity, which currently are measured manually.

A variety of automatic methods have been developed for segmenting kidneys in 2D US images, including active contour model (ACM) based methods (5, 6), atlas-based methods (7), Markov random field based methods (8), watershed based methods (9), machine learning based methods (10), and deep learning methods (11, 12). Among them, the ACM based method is appealing for its robustness to imaging noise, weak boundaries, and large appearance variation within kidneys. However, the existing ACM based image segmentation methods typically adopt gradient descent flow based optimization techniques, which often get stuck at local minima (13). Such a limitation can be overcome by graph cuts (GC) techniques (13–15). Particularly, GC techniques model the image segmentation task as an image labeling problem on a graph (16–20). The GC techniques can be integrated with the ACM based methods by transforming the minimization problem of the ACM methods into a min-cut problem of a graph (13–15, 19, 21, 22). However, speckle noise and low contrast of US images might degrade the segmentation performance if the US images are directly used as input to the segmentation algorithms.

Texture analysis has been adopted to characterize US images in image segmentation studies and led to better performance than image segmentation based on image intensity information alone (23–26). Texture feature extraction techniques can be categorized as structural (27–29), statistical (30), model-based (23, 31), and transform-based methods (25, 26, 32–35). In particular, structural methods utilize geometric primitives to represent textures, statistical methods represent texture features by statistics of intensity distributions or relationships among them, and model-based methods generally assume that texture features obey a certain statistical distribution. Different from aforementioned methods, transform-based methods are typically built upon multi-scale frequency and multi-direction analysis and have been demonstrated capable of capturing texture information more effectively (33). Particularly, Gabor transformation (26, 33, 36) is a representative transform based method for characterizing image texture information. A Gabor transformation consists of a set of filters, which are multiplications of a Gaussian function and a sinusoidal function in different directions and frequencies, capable of capturing texture information similar to perception in the human visual system (33, 37, 38). Although many US image segmentation studies have adopted texture information in the image segmentation (23–26), most of them only consider a single texture feature map for image segmentation, which might be insufficient to deal with imaging noise of clinical data.

Building upon the success of existing image segmentation methods (13, 14, 16, 17, 19), we propose a new graph cuts based method to segment kidney US images by integrating original image intensity information and texture feature maps extracted using Gabor filters. To handle large appearance variation within kidney images and improve computational efficiency, we build a graph of image pixels close to kidney boundary instead of building a graph of the whole image. To make the kidney segmentation robust to weak boundaries, we adopt localized regional information to measure similarity between image pixels for computing edge weights to build the graph of image pixels. The localized graph is dynamically updated and the GC based segmentation iteratively progresses until convergence. Our method has been evaluated based on clinical kidney US images of 85 subjects.

MATERIALS AND METHODS

Our image segmentation method is built upon graph cuts and texture analysis techniques (16, 17, 33). The image segmentation problem is solved using GC techniques by building a dynamically evolving graph of image pixels close to the boundary of kidneys. Both original image intensity information and multi-scale and multi-directional Gabor features are integrated into a localized regional similarity measure for measuring similarity between image pixels.

Texture feature map extraction

A new method is proposed to extract texture feature maps based on multi-scale and multi-direction analysis using Gabor transformation.

Gabor transform is a windowed Fourier transform (33), consisting of Gabor filters in different scales and directions. A 2D Gabor filter is the multiplication of a Gaussian kernel function and a sine wave function:

$$\{\begin{array}{c}g=\frac{1}{2\pi {\sigma }_x{\sigma }_y}\mathit{\text{exp }}(-(\frac{x\prime }{{\sigma }_x^2}+\frac{y\prime }{{\sigma }_y^2}\left)\right)\mathit{\text{ exp }}\left(i\right(2\pi \frac{x\prime }{\lambda }\left)\right),\hfill \\ x\prime =x\mathit{\text{cos}}\theta +y\mathit{\text{sin}}\theta ,y\prime =-x\mathit{\text{sin}}\theta +y\mathit{\text{cos}}\theta \hfill \end{array}$$ (1)

where σ_x and σ_y are standard deviations of the Gaussian functions along axis x and axis y respectively, θ is the direction of the Gabor filter, wavelength λ is the frequency factor, and the frequency $\omega =\frac{2\pi }{\lambda }$. A Gabor filter with a small λ captures image information of high frequency and small scale, while image information of low frequency and big scale can be captured by a filter with a big λ. Gabor filters in different scales and directions can be integrated to capture multiscale image information (39).

Given an US image I, using Gabor filters we compute texture features F_i,j, i = 1, …, m, j = 1, …, n at m scales in n directions with values f_i,j(p) at pixel p of I. To enhance kidney boundaries and suppress noise in US images, we propose a novel scheme to fuse filtered images with greater responses, which amounts to a filtering process along edges in US images, as schematically illustrated in Fig. 1. Particularly, given a pixel p with filtered values f_i,j(p) corresponding to i = 1, …, m scales and j = 1, …, n directions, a dominant direction at each scale is first determined as

$$g_i={\text{argmax}}_{j=1,\dots ,n}|f_{i,j}(p)|,i=1,\dots ,m,$$ (2)

where the dominant direction at scale i has the maximum absolute value, |f_i,gi(p)|, among all directions. For every direction, we then compute the number of scales at which the direction under consideration is the dominant direction.

$$D_{\mathrm{j}}={\sum }_{i=1}^m\delta (|f_{i,j}(p)|-|f_{i,g_i}(p)|),\mathrm{j}=1,\dots ,n$$ (3)

where function δ(·) is 1 at zero and 0 otherwise. Finally, we identify dominant directions across all scales as

$$\mathrm{\Omega }=\{j|D_{\mathrm{j}}\ge 0.5·{\text{max}}_{j=1,\dots ,n}{D}_{\mathrm{j}}\},$$ (4)

where dominant directions are identified as those with filtering responses greater than half of the maximal response values to balance robustness and redundancy of the dominant directions.

Fig. 1.

Feature map extraction. Filtered images with greater responses at different scales and in different directions are fused to extract texture feature maps based on the Gabor transform.

After determining the multiple dominant directions (for instance, d_j1, d_j2, d_j3 as shown in Fig. 1), we compute a texture feature map F with element f(p) by fusing the filtered images in all dominant directions across all scales

$$f(p)=\frac{1}{m}{\sum }_{i=1}^m{\sum }_{j\in \mathrm{\Omega }}|f_{i,j}(p)|.$$ (5)

Example Gabor feature maps are shown in Fig. 2, along with their corresponding US images of 3 subjects. Both the US images and their corresponding Gabor features are to be used as feature maps in our image segmentation algorithm.

Fig. 2.

Example feature maps of kidney US images. Top row: Original images; bottom row: Gabor feature maps.

Dynamic GC based image segmentation with integrated multiple feature maps

The GC based image segmentation methods model the image segmentation as a graph partition problem (40). Given a set of pixels P to be segmented into a set of regions, the GC based segmentation methods model the pixels P as graph nodes V that are connected by edges E weighted by image similarity measures W, i.e., the image is modeled as a graph G = (V, E, W). A cut on the graph G can be seen as a contour on the graph, which separates graph nodes V into two disjoint subsets (segments) S and T, and the cost of the cut is measured by |CUT_C|_G = Σ_p∈S,q∈T w(p, q), where w(p, q) is a weight of the edge connecting pixels p and q. The cut with the minimal cost (min-cut) can be obtained using min-cut/max-flow algorithms (16).

The performance of GC based image segmentation is hinged on the graph of pixels to be segmented, particularly the weights of edges that connect the pixels. Since the kidney has large appearance variations and weak boundaries in US images, it is difficult to build a graph of all pixels of an US image to obtain satisfied segmentation performance. Motivated by the success of localized image segmentation method (14, 19), we build a graph of pixels surrounding the kidney boundaries and gauge similarity between pixels using a combination of pixel and regional measures based on the multiple feature maps, as illustrated in Fig. 3. Particularly, a narrow band surrounding the kidney boundaries is identified by inflating and shrinking an initialization of kidney boundary with radii r₁ and r₂ respectively. The weight w(p, q) for an edge connecting two pixels p and q is defined as an image similarity measure between the pixels:

$$w(p,q)=\alpha ·w_p(p,q)+(1-\alpha )·w_r(p,q),$$

$$w_p(p,q)=\mathit{\text{exp }}(-\frac{\frac{1}{N}{\sum }_{i=1}^N{(I_i(p)-I_i(q))}^2}{\sigma }),w_r(p,q)=\mathit{\text{exp }}(-\frac{(1/(\frac{1}{N}{\sum }_{i=1}^N{K}_i(p,q)))}{\sigma }),$$ (6)

where p and q are adjacent pixels, w_p(p, q) and w_r(p, q) are image similarity measures based on pixel information and regional information respectively, α is a trade-off parameter between the two terms, I_i(p) and I_i(q) are image intensity value of p and q for the ith feature map, K_i(p, q) is an image difference measure based on regional information, σ is a parameter, and i = 1, 2 corresponding to the original US image intensity and the Gabor feature map respectively. The numerator terms in both w_p(p, q) and w_r(p, q) are normalized to [0,1]. Particularly, K_i(p, q) is defined as

$$K_i(p,q)={\text{min}}_{(l_p,l_q)\in \{(S,T),(T,S),(S,S)\mathit{\text{ or }}(T,T)\}}{(I_i(p)-f_i^{l_p}(p))}^2+{(I_i(q)-f_i^{l_q}(q))}^2,$$ (7)

where I_i(p) and I_i(q) are image intensity value of p and q for the ith feature map, $f_i^S(p),f_i^T(p),f_i^S(q),f_i^T(q)$, are intensity mean of pixels of a small neighborhood of p and q corresponding to segments S and T respectively as illustrated in Fig. 3(c), and l_p ∈ {S, T} and l_q ∈ {S, T} are possible segmentation labels of p and q to be updated according to the minimal value computed by Eq. (7). Particularly, K_i(p, q) measures differences between image intensity values of pixels and mean image intensity values of their corresponding optimal segmentation results. However, if the optimal segmentation labels of p and q are the same, we set K_i(p, q) = max_{(u,v)∈(V,V)} K_i(u, v) so that they will not be separated into different segments.

Fig. 3.

A graph of pixels within a narrow band of the kidney boundary. (a) A narrow band surrounding the kidney boundary: the green curve is an initialization of kidney boundary and a narrow band is located in-between the blue curves which are obtained by inflating and shrinking the green contour; (b) A graph of pixels within the narrow band (green pixels) in-between the blue cures (denoted by yellow pixels): n-links connect neighboring pixels (p, q), and t-links connect yellow pixels inside or outside the green pixels with S or T respectively so that the pixels are segmented into disjoint subsets of S or T; (c) Illustration of the computation of $f_i^S(p)$ and $f_i^T(p)$ given a segmentation result with segments S and T.

In the present study, a circular neighborhood system with radius r is adopted for computing intensity mean of pixels of a small neighborhood. As illustrated in Fig. 3(c), $f_i^S(p)$ and $f_i^T(p)$ of a pixel p are intensity means of pixels within segments S and T respectively.

The edge weights and the segmentation result in the graph are updated iteratively until convergence. The segmentation algorithm is summarized as:

1. Compute Gabor feature maps given US images and the numbers of scales and directions.

2. Initialize a contour C around but inside kidney boundary, and assign pixels with labels: $\mathit{\text{BW}}(p)=\{\begin{array}{c}1,\text{ if }p\text{ is inside the contour},\hfill \\ 0,\text{ otherwise}.\hfill \end{array}$

3. Build a narrow band of the contour C by inflating and shrinking the initialization contour as illustrated in Fig. 3(a) with radii r₁ and r₂, respectively.

4. Build the graph of pixels within the constructed narrow band as illustrated in Fig.3 (b):

   n-links: pixels within the narrow band are connected with n-links weighted with w(p, q) = w_p(p, q) + w_r(p, q) according to Eq. (6);

   t-links: pixels on the boundaries of the narrow band are connected with segments S and T weighted with ∞.

5. Apply max-flow/min-cut algorithm (16) to the graph, and assign pixels with labels: $\mathit{\text{BW}}(p)=\{\begin{array}{c}1,\text{ if }p\text{ is inside the contour},\hfill \\ 0,\text{ otherwise}.\hfill \end{array}$

6. Update the regional measure similarity w_r(p, q) according the result contour.

7. Repeat steps (4)–(6) until converge.

Segmentation performance evaluation and parameter optimization

The proposed method has been validated based on 2D kidney US images obtained from 85 subjects with their original image intensity information and Gabor feature maps as image features. The 2D US images were collected using a Philips IU22 Ultrasound system at a frequency of 55HZ with a pediatric abdominal transducer, and all identification information was removed. The work described has been carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments. The study has been reviewed and approved by the IRB.

Each subject’s bilateral kidney images were manually segmented and treated as ground-truth for evaluating the automatic image segmentation performance. For evaluating the image segmentation performance, we adopted Dice Index, Jaccard Index, and Mean Distance to quantitatively measure similarity/difference between the automatic segmentation results and manual labels. These quantitative metrics are defined as $\text{Dice Index}=2\frac{V(E\cap F)}{V(E)+V(F)}$, Mean Distance = mean_e∈BE (min_f∈BFd(e, f)), $\text{Specificity}=\frac{V(\overline{E}\cap \overline{F})}{V(\overline{E}\cap \overline{F})+V(E\cap \overline{F})},\text{ Accuracy}=\frac{V(E\cap F)+V(\overline{E}\cap \overline{F})}{V(E\cap F)+V(\overline{E}\cap \overline{F})+V(E\cap \overline{F})+V(\overline{E}\cap F)}$ where E is an automatic segmentation result, F is a manual segmentation result, Ē and F̅ are the complementary sets of E and F respectively, d(·,·) is Euclidian distance between two points, and BE and BF are boundary voxels of E and F respectively.

We randomly selected US images of 20 subjects to optimize parameters of the proposed method and evaluate how the parameters affect the segmentation performance. Based on the optimization parameters, we validated the algorithm based on US images of the remaining 65 subjects. We compared the proposed method based on 2D US images of 65 subjects with the image segmentation methods, including localizing region based active contours (LRGC) (41) and geodesic active contours (GAC) (42). In particular, the implementation of LRGC and GAC was based on a software package “creaseg” (43) that integrated 6 different image segmentation methods (41, 42, 44–47), and LRGC and GAC methods were among the best. Moreover, a recent adaptive multi-feature segmentation model (AMFSM) (48) was also adopted in comparison. The same initialization contours were used by all the methods under comparison. All the experiments were implemented using MATLAB R2014b on a Desktop PC with 64-bit Windows operating system (i5-3470 CPU, 3.2 GHz, Quad core, 8G RAM). On average, it took ~16 seconds for our method to segment one kidney image, including ~13 seconds for the feature extraction and ~3 seconds for the segmentation.

RESULTS

Optimization of the parameters

The proposed method has several parameters, including the numbers of scales and directions for computing the Gabor feature maps, σ for computing the image similarity measures, α for balancing the pixel and regional information terms, radius r for the circular neighborhood system, parameters r₁ and r₂ for building the narrow band by inflating and shrinking operations. A grid searching was adopted to select an optimal setting for the parameters based on US images of 20 subjects. In particular, the number of scales was selected from {2,3,4,5}, the number of directions was selected from {4,8,16}, σ was selected from {1, 0.1, 0.01}, α was selected from {0, 0.25, 0.5, 0.75, 1}, radius r was selected from {5,10,15}, parameters (r₁, r₂) were selected from {(15, 3), (20, 3), (10, 3), (15, 1), (15, 5)}. Three different initialization kidney contours were adopted to segment each kidney,, and the parameter optimization was determined based on the average segmentation performance f 3 segmentation results. The experiment results revealed that the best kidney segmentation performance (Dice Index=0.9594) on the training data could be obtained with the number of scales=4, the number of directions=8, σ = 0.1, α = 0.75, r = 5, (r₁, r₂) = (15, 1). This setting was adopted in all following experiments unless otherwise specified.

Experiments on single/multiple feature maps

Our method has been evaluated with different setting of image features, including segmenting images based on the original images, the Gabor feature map, and their combination based on US images of 20 subjects. The kidney segmentation based on either the original image intensity information or the Gabor feature map alone is a degraded version of our proposed method. Box plots of segmentation accuracy measures of single/multiple feature maps and segmentation results of an example US image are shown in Fig. 4. These results demonstrated that the proposed multiple feature maps based method could achieve better performance than the alternatives. Therefore, the setting of multiple feature maps was adopted in all following experiments.

Fig. 4.

Kidney segmentation result on US images. Top row: Red curves are segmentation results obtained by the automatic segmentation algorithm with different feature maps, and yellow curves are manual segmentation results; (a1) Segmentation result obtained based on the original image intensity information; (a2) Segmentation result obtained based on the Gabor feature maps; (a3) Segmentation results obtained based on their combination. Bottom row: Box plots of kidney (left and right) image segmentation accuracy measures of 20 subjects based on the original image intensity, the Gabor feature map, and their combination. (b1–b4) are Dice, Mean Distance, Specificity, and Accuracy, respectively.

Experiments on different settings of image similarity measures

The image similarity measure was defined based on the combination of both pixel information and local regional information. We evaluated how the image similarity measure affected the image segmentation performance. Example segmentation results of one subject obtained with different settings α = {0, 0.25, 0.5, 0.75, 1} are shown in the top row of Fig. 5 (a1–a5), and the bottom row of Fig. 5 (b1–b4) shows segmentation accuracy measures of US images of 20 subjects obtained with the image similarity measures defined based on the pixel information, the regional information, and their different combinations weighted by parameter α. These experiments demonstrated that the image similarity measure defined based on the combination of both the pixel information and the regional information with α = 0.75 yielded the best kidney segmentation performance.

Fig. 5.

Comparison experiments utilizing only pixel information, only regional information, and their different combinations. The top row shows segmentation results of one subject based on only pixel information (a1), only regional information (a5), and their combinations (a2–a4 with α = 0.25, 0.5, 0.75). The bottom row shows box plots of segmentation performance measures of 20 subjects. (b1–b4) are Dice, Mean Distance, Specificity, and Accuracy, respectively, with α = 0, 0.25, 0.5, 0.75, 1 for comparison, and α = 0.75 is utilized in our method.

Experiments on different initializations

The proposed method needs to have an initialization contour which could be obtained by manually picking 6–10 points to outline the shape of the kidneys. The top row of Fig. 6 (a1–a3) shows example initialization contours for 3 US images. To evaluate how the initialization affects the segmentation performance, we obtained segmentation results of US images of the 20 subjects with 3 different initialization contours. Quantitative segmentation results as shown in the bottom row of Fig. 6 (b1–b4) indicated that the segmentation performance was relatively stable with respect to the initialization contours. The degree of consistency of the segmentation performance measures obtained with different initializations was further measured by intraclass correlation coefficients (ICCs, the segmentation performance measures was modeled as averages of k independent measurements on randomly selected objects) (49), as summarized in Table 1. The ICC results indicated that the proposed method was relatively robust to different initializations.

Fig. 6.

Experimental results of different initializations. The top row: 3 initialization examples in (a1–a3). Red curves are initialization contours determined by the points in green. The bottom row: Box plots of kidney image segmentation results for 20 subjects with 3 different initializations. (b1–b4) are Dice, Mean Distance, Specificity, and Accuracy, respectively, with 1,2,3 for 3 initializations.

Table 1.

Intraclass correlation coefficients for 3 different initializations

Dice Index	Mean Distance	Specificity	Accuracy
0.8531	0.8329	0.8549	0.8691

Experiments on different narrow bands

Example segmentation results of one subject obtained with different narrow bands constructed with settings (r₁, r₂) ∈ {(15, 3), (20, 3), (10, 3), (15, 1), (15, 5)} are shown in the top row of Fig. 7 (a1–a5), and the bottom row of Fig. 7 (b1–b4) shows segmentation accuracy measures of US images of 20 subjects. The experiments demonstrated that the setting (r₁, r₂) = (15, 1) of obtained the best kidney segmentation performance.

Fig. 7.

Experimental results of different inflating and shrinking parameters r₁ and r₂ for building narrow band. The top row: 5 parameter settings experiments in (a1–a5) with r₁/r₂ ∈ {15/3, 20/3, 10/3, 15/1, 15/5}. The bottom row: Box plots of kidney image segmentation results for 20 subjects with the 5 parameter settings in (b1–b4).

Comparison with alternative methods

Fig. 8 shows box plots of segmentation performance measures of results obtained by different methods under comparison, and the quantitative segmentation performance measures are summarized in Table 3, indicating that the proposed method performed significantly better than the alternatives (p<0.05, paired Wilcoxon rank sum tests). The experiment results demonstrated that the proposed method achieved better performance than others.

Fig. 8.

Segmentation performance measures of segmentation on 2D US images of 65 subjects obtained by different methods. (a1–a4) are Dice, Mean Distance, Specificity, and Accuracy, respectively.

DISCUSSION AND CONCLUSIONS

In this paper, a dynamic GC based segmentation method with integrated multiple feature maps is proposed to segment kidneys in 2D US images. In particular, the proposed method consists of texture feature extraction and multi-channel GC based segmentation. The feature maps used in our method include the original kidney US image intensity information and the Gabor feature maps. The proposed method has been evaluated on US images of 65 subjects with parameters optimized based on US images of 20 subjects different from the validation dataset.

We have also compared our methods with the image segmentation methods, including localizing region based active contours (41), geodesic active contours (42), and a recent adaptive multi-feature segmentation model (48). Extensive experiment results have demonstrated that the combination of multiple feature maps can yield better segmentation performance than any single feature map. The proposed method will facilitate development of accurate and reproducible ways to objectively measure kidney features such as echogenicity and cortico-medullary differentiation, which are increased and decreased, respectively, in many kidney diseases. Development of reproducible method to objectively measure these features will lead to identification and validation of anatomic biomarkers that may predict clinically important outcomes such as progression of chronic kidney disease.

The image similarity measure plays a critical role in the GC based segmentation methods. To deal with high speckle noise and low contrast between foreground and background, as well as weak boundaries and large appearance variations in kidney US images, the proposed method integrates multi-feature maps into pixel and localized regional similarity measures to build a narrow band graph. The experimental results have demonstrated that the multi-feature maps based method could yield better performance than any single feature based segmentation.

The proposed method is a semi-automatic segmentation method, and manual initialization is needed. Although the high speckle noise and low contrast in kidney US images may result in instable segmentation results if different initializations are used, good performance can be obtained by constraining the initialization around the kidney boundary and building a narrow band graph. The ICC results also indicated that the segmentation results of the proposed method were robust to different initializations.

We compared our method with the image segmentation methods, including localizing region based active contours (41), geodesic active contours (42), and a recent adaptive multi-feature segmentation model (48). The former two methods are the typical image segmentation methods built upon regional and pixelwise information respectively, and have the best performance among the six methods in the “creaseg” software (43). The AMFSM (48) is capable of integrating multiple features using a strategy different from ours. In particular, the AMFSM takes a cosine similarity measure of feature vectors to determine the curve evolution direction and a distance similarity measure to determine the magnitude of driving force, while our method utilize the GC based method to determine the final results by integrating multi-feature maps together and building a narrow band graph, which tends to achieve a global optimal solution. Based on the same image features, our method achieved better segmentation performance than the AMFSM, indicating that our strategy for integrating multiple features is more suitable for Kidney US image segmentation.

Deep learning techniques have been widely adopted in image segmentation studies. A recent study has reported that a deep learning based method obtained an average Dice index of 0.8395 for segmenting kidneys in 2D Ultrasound images (11). The inferior performance of deep learning methods might be due to large appearance variations of the kidneys in 2D US images. In contrast, our method is a semi-automatic algorithm and could achieve good segmentation performance with a reasonable initialization.

In summary, we propose a dynamic localized GC based segmentation method with integrated multiple feature maps to improve the segmentation of Kidney US images. Besides the original image intensity information, Gabor texture feature map is extracted from US images and used as a complementary feature map. The multi-feature maps are finally fused via a novel dynamic localized multi-channel GC. The experimental results demonstrated that the proposed method could achieve promising segmentation performance for segmenting kidneys in US images. Our method can be integrated with US image interpretation systems to automatically compute anatomical measures of kidneys that could be informative for predicting risk of end-stage renal disease (3).

Table 2.

Segmentation accuracy measures of results obtained by different methods (Paired Wilcoxon rank sum tests were adopted to compare the proposed method with the alternatives)

	Dice Index				MeanDistance
	mean	std	median	p-value	mean	std	median	p-value
LRAC	0.9051	0.0257	0.9093	5.12e-23	3.6527	0.9810	3.5612	8.14e-23
GAC	0.9315	0.0374	0.9465	3.99e-04	2.9888	2.0283	2.2896	1.02e-04
AMFSM	0.9024	0.0388	0.9118	2.04e-22	4.1275	1.1435	3.9175	1.51e-22
Proposed	0.9446	0.0179	0.9489		2.2551	0.5783	2.1941
	Specificity				Accuracy
	mean	std	median	p-value	mean	std	median	p-value
LRAC	0.9932	0.0048	0.9941	3.73e-13	0.9868	0.0053	0.9877	1.02e-22
GAC	0.9945	0.0087	0.9969	1.24e-04	0.9894	0.0085	0.9918	1.00e-03
AMFSM	0.9901	0.0065	0.9915	6.33e-21	0.9854	0.0054	0.9858	1.75e-22
Proposed	0.9971	0.0022	0.9976		0.9919	0.0031	0.9924

List of abbreviations

US

Ultrasound

ACM

active contour model

GC

graph cuts

LRGC

localizing region based active contours

GAC

geodesic active contours

AMFSM

adaptive multi-feature segmentation model

ICCs

intraclass correlation coefficients