Pattern Recognition Letters 34 (2013) 1470–1475

Contents lists available at SciVerse ScienceDirect

Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec

Kidney segmentation using graph cuts and pixel connectivity Ashish K. Rudra a, Ananda S. Chowdhury a,⇑, Ahmed Elnakib b, Fahmi Khalifa b, Ahmed Soliman b, Garth Beache c, Ayman El-Baz b a

Department of Electronics and Telecommunication Engineering, Jadavpur University, Kolkata, India BioImaging Laboratory, Bioengineering Department, University of Louisville, Louisville, KY, USA c Department of Diagnostic Radiology, School of Medicine, University of Louisville, Louisville, KY, USA b

a r t i c l e

i n f o

Article history: Received 23 April 2012 Available online 23 May 2013 Communicated by A. Shokoufandeh Keywords: Kidney segmentation Graph cut Connectivity analysis Pixel labeling Dijkstra’s shortest path algorithm

a b s t r a c t Kidney segmentation from abdominal MRI data is used as an effective and accurate indicator for renal function in many clinical situations. The goal of this research is to accurately segment kidney from very low contrast MRI data. The present problem becomes challenging mainly due to poor contrast, high noise and partial volume effects introduced during the scanning process. In this paper, we propose a novel kidney segmentation algorithm using graph cuts and pixel connectivity. A connectivity term is introduced in the energy function of the standard graph cut via pixel labeling. Each pixel is assigned a different label based on its probabilities to belong to two different segmentation classes and probabilities of its neighbors to belong to these segmentation classes. The labeling process is formulated according to Dijkstra’s shortest path algorithm. Experimental results yield a (mean ± s.d.) Dice coefficient value of (98.60 ± 0.52)% on 25 datasets. Ó 2013 Elsevier B.V. All rights reserved.

1. Introduction Acute rejection – the immunological response of the human immune system to the foreign kidney is the most important cause of graft failure after renal transplantation (Rigg, 1995). Currently, the diagnosis of rejection is done via biopsy. However, biopsy has the downside effect of subjecting the patients to risks like bleeding and infections (Yang et al., 2001). Therefore, a noninvasive and repeatable technique like computer-aided segmentation of kidney becomes essential for the diagnosis of acute renal rejection. From the accurate segmentation, agent kinetic curves (i.e., signal intensity versus time curves, or perfusion curves) are constructed. From these curves, perfusion-related indexes (e.g., peak signal intensity, time-to-peak, and initial up-slope) are estimated and are used to distinguish between rejection groups and non-rejection groups (Khalifa et al., 2010). Our input for this problem is dynamic MRI data obtained using gradient-echo T1 imaging with a 1.5T MRI scanner (Signa Horizon LX Echo speed; General Electric Medical Systems, Milwaukee, WI, USA). In particular, some images of the time series data have extremely poor contrast. This is the reason why we first focus on kidney segmentation in 2D slices and later show the extension on entire 3D volumes. The main factors which make this problem of kidney segmentation very difficult in addi-

⇑ Corresponding author. Tel.: +91 33 2457 2405; fax: +91 33 2414 6217. E-mail address: [email protected] (A.S. Chowdhury). 0167-8655/$ - see front matter Ó 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.patrec.2013.05.013

tion to extremely poor contrast are unclear borders between the kidney and its background, image acquisition artifacts, image noise, and various pathologies, such as tumors and nephrolithiasis. Lin et al. (2011) have proposed a solution for fully automatic kidney segmentation. Temporal changes in intensity counts, intensity-pair distribution image contrast enhancement method, adaptive thresholding, and morphological operations are applied for that purpose. However, this work has several processing stages. In this work, we focus on improving graph cuts, a single-step process, for kidney segmentation. It is a well-known fact that the graph cut-based segmentation method provides globally optimal segmentation results (Boykov and Funka-Lea, 2006; Boykov and Jolly, 2001). Semi-automated kidney segmentation using graph cuts is reported in Hackjoon et al. (2009), Ali et al. (2007). A variety of methods have been proposed to incorporate fixed parametric shape information to spatially constrain the graph min-cut optimization to increase the accuracy of the segmentation. For example, see the works of Slabaugh and Unal (2005), and Freedman and Zhang (2005). Yuksel et al. (2006) use the Poisson distribution and distance maps to compute the shape term of the graph for the segmentation of 2D kidney slices from DCE-MRI. The parametric methods (Ali et al., 2007; Slabaugh and Unal, 2005; Freedman and Zhang, 2005; Yuksel et al., 2006; Malcolm et al., 2007) require a specific shape model which is found to be inadequate for kidney segmentation due to the large inter-patient shape variability. In addition, an interactive initialization is often required to register the shape model with the image. Use of generic shape information

1471

A.K. Rudra et al. / Pattern Recognition Letters 34 (2013) 1470–1475

in the graph cut framework has emerged as another important strategy which obviates the need of computationally intensive registration of shape model and the image. Veksler (2008) used a generic star-shaped prior in graph cut-based segmentation framework. But a disadvantage of Veksler (2008) is that only a shape obeying generic star shape can be extracted. As a statistical alternative, non-parametric shape modeling techniques are introduced (Sabuncu et al., 2009). Freiman et al. (2010) employ an iterative non-parametric model-based graph min-cut approach for kidney segmentation in CT images. There are also several temporal kidney segmentation approaches available in the literature (Chevaillier et al., 2008; Zöllner et al., 2009; Beatrice et al., 2009). Note that the non-parametric as well the temporal methods are computationally quite intensive. So, we notice that both parametric and non-parametric approaches to shape modeling have certain limitations. Moreover, none of the reported methods explicitly address the problem of extremely poor contrast which is the most important concern of this problem. In an earlier work (Rudra et al., 2011), we have shown how the Boykov–Jolly energy function can be modified to handle the problem of low contrast. However, we find that even this method fails to produce very high quality segmentation results for extremely low contrast images. So, we improve the algorithm described in Rudra et al. (2011) by incorporating the connectivity analysis of the image pixels following the Dijkstra’s shortest path algorithm. Note that graph cut provides an efficient global optimization framework where one can integrate region, boundary and connectivity information. Graph cuts using connectivity information can be found in the works of Vicente et al. (2008), where three connectivity constraints are imposed in the segmentation framework. In the proposed work, we use connectivity in an altogether different manner by classifying the pixels into three different groups. The main advantages of our method are as follows: (i) no shape prior is needed, (ii) no additional user interaction is required to perform the connectivity analysis, and (iii) the method can segment very low contrast kidney images with high accuracy in both 2D and 3D. The rest of the paper is organized in the following manner: in Section 2, we describe the method in details along with necessary theoretical foundations. In Section 3, we present and analyze the experimental results with necessary comparisons. In Section 4 we conclude the paper and mention the directions for future research. 2. Method An image is modeled as a weighted undirected graph G = G (V, E). Let P denote the set of all pixels. Each pixel p 2 P constitutes a node/vertex in G. In addition, two special terminal nodes, namely, the ‘source’(s) and the ‘sink’ (t) are considered following Ford and Fulkerson, (1962). Ne(p) denotes the neighborhood of pixel p. The graph G contains two types of edges: the neighborhood links/nlinks (N) and the terminal links/t-links (T). Each node p has two t-links {s, p} and {p, t}, connecting it to the source and the sink. The n-links are constructed between each pixel p and its neighboring pixels q (q 2 Ne(p)). We use |Ne(p)| = 8. A new type of edge called connectivity links/c-links (C) is introduced. The c-links connect every pixel p to the user specified pixel u, which can be any object seed. The weight of c-link of a pixel depends on how well the pixel is connected to the node u. For this graph G, we can write:

V ¼P[s[t

ð1aÞ

E¼T [N[C

ð1bÞ

A. Graph cut for low contrast images Let A define segmentation, i.e., classification of all pixels into either ‘‘object’’ or ‘‘background’’. The corresponding energy function, to be optimized, can be written as:

EðAÞ ¼ BoðAÞ þ kRðAÞ þ lCðAÞ

ð2Þ

where the boundary properties are given by the term Bo(A), the region properties are given by the term R(A), the connectivity properties are given by the term C(A), and k and l are weighting factors. For a low contrast image, Bo(A) and R(A) can be mathematically expressed as (Rudra et al., 2011):

X

BoðAÞ ¼

Bfp;qg

ð3Þ

p;q2NeðpÞ

RðAÞ ¼

X Rp ðAp Þ

ð4Þ

p2P

where,

Bofp;qg ¼ K fp;qg expððIp  Iq Þ2 =2r2 Þ 

1 dðp; qÞ

ð5Þ

2 Rp ð\obj"Þ ¼ K obj p expðððSA  SB Þ=DÞ Þ lnðPrðI p jobjÞÞ

ð6Þ

2 Rp ð\bkg"Þ ¼ K bkg p expðððSA  SB Þ=DÞ Þ lnðPrðIp jbkgÞÞ

ð7Þ

Pr(Ip|obj) and Pr(Ip|bkg) are obtained from the object and background seeds inputted by the user (Boykov and Funka-Lea, 2006; Boykov and Jolly, 2001). Expressions for SA, SB and D are given below:

SA ¼

jAq j X PrðIqjobjÞ=distðp; qi Þ

ð8Þ

i¼1

SB ¼

jBq j X PrðIqjbkgÞ=distðp; qi Þ

ð9Þ

i¼1

D¼

jNeðpÞj X

ð1=distðp; qi ÞÞ

ð10Þ

i¼1

The coefficient K{p,q} in Eq. (5) is evaluated by comparing the probabilities of any two voxels p 2 P and q 2 NeðpÞ to belong to the same segmentation class using the terms Pr(Ip|obj), Pr(Ip|bkg), Pr(Iq|obj) and Pr(Iq|bkg). The coefficient K obj p in Eq. (6) and the coefficient K bkg in Eq. (7) are evaluated on the basis of the probability of p a single voxel p to belong to the segmentation class ‘‘object’’ or ‘‘background’’ using the terms Pr(Ip|obj) and Pr(Ip|bkg). For the detailed algorithm which has been used to evaluate K{p,q},K obj and p K bkg p , see Rudra et al. (2011). The third term C(A) in Eq. (2) can be expressed as:

CðAÞ ¼

X C p ðAp Þ

ð11Þ

p2P

The value of CP depends on how well the node p is connected to the user specified node u. The strength of the connection between u and p is determined according to the characteristics of the shortest path (Wp) between these nodes. The shortest path is obtained following Dijkstra’s shortest path algorithm with an appropriate grammar. Note that this algorithm is applied on a graph G1 = G1(V1, E1), different from the graph G. For the graph G1, we can write:

V1 ¼ P

ð12aÞ

E1 ¼ N

ð12bÞ

B. Pixel connectivity analysis We first introduce some definitions to explain the concept of connectivity.

1472

A.K. Rudra et al. / Pattern Recognition Letters 34 (2013) 1470–1475

Definition 1 (Class A node). A node p is termed as class A if it has a higher probability to belong to the ‘‘object’’ class, compared to that of the ‘‘background’’ class.

Definition 2 (Class B node). A node p is termed as class B if it has higher probability to belong to the ‘‘background’’ class, compared to that of the ‘‘object’’ class. Definition 3 (L1 node). A node p is labeled L1 if it is a class A node and all intervening nodes on the shortest path Wp from u to p are class A nodes. Definition 4 (L2 node). A node p is labeled L2 if any of the following conditions hold: (i) All the intervening nodes on the shortest path Wp from u to p are class A nodes but node p is a class B node. (ii) Node p is a class A node and only one of the intervening nodes of Wp is a class B node

Definition 5 (L3 node). A node p is labeled L3 node if any of the following conditions hold: (i) Node p is a class B node and only one of the intervening nodes of Wp is a class B node. (ii) More than one transit nodes on Wp are class B nodes. Let us assume that an user specifies a node u within the object boundary. So, the node u is essentially a class A node (vide Definition 1) and it can be easily concluded (vide Definition 3) that its label is L1. On the basis of the nature of shortest path between u and any other node p, we label all the nodes as L1 or L2 or L3. We show the structure of the graph G we have used in this paper in Fig.1. Now, we state and prove two important theorems in the context of labeling of nodes.

to p, all transit nodes should be of type class A. Hence, p is not an L1 node. h Note that we can determine the label of a node, if the labels of all the neighbors and its own class are known to us. Thus, the problem of labeling all the nodes can be potentially solved by a dynamic programming approach, similar to the Dijkstra’s shortest path algorithm (Cormen et al., 2001). The labeling process essentially captures the object shape model in a unique manner. L1 nodes have highest affinity and hence are strongly connected with the object seed u. L3 nodes have lowest affinity and hence are weakly connected with the object seed u. L2 nodes have moderate affinity and hence are moderately (less strongly than L1 but more strongly than L2) connected with the object seed u. If we draw the image in which the intensity of the pixels is proportional to the label of the corresponding node, it can be seen that the L1 nodes form the kidney (object), L3 nodes form the surrounding regions (background) and L2 nodes lie on the boundary of background and foreground regions (edge). In the next paragraph, we present an algorithm which determines the shortest paths from node u to all other pixels p in the image and label these pixels dynamically. A path (between the user specified node u and any node p in the image) is denoted by (nL_3, nL_2, nL_1) where nL_3, nL_2, nL_1 denote number of L3, L2, L1 intervening nodes on the path, respectively. Grammar: Path i is shorter than the path j [(niL 3 ; niL 2 ; niL 1 ,) < (njL 3 ; njL 2 ; njL 1 )] if any of the three conditions hold: (i) niL (ii) niL (iii) niL

3 3 3

< njL = njL = njL

3 3 3

and niL and niL

2 2

< njL = njL

2 2

and niL

1

< njL

1

Variables: P = set of all nodes u = user specified node Lp = (0, 0, 1) if the node is an L1 node (0, 1, 0) if the node is an L2 node (1, 0, 0) if the node is an L3 node W ip = the shortest/best path from u to p, at stage i of the algorithm. Wp = the final shortest/best path to node p T1 = set of the nodes, so far computed by the algorithm

Theorem 1. For an L1 node, at least one of its 8 neighbors must also be an L1 node. Step 1. Initialization: T1 = {u} Proof. Let p be an L1 node. Then, all the intermediate nodes along the shortest path Wp (from u to p) are class A nodes. To determine a shortest path from any node, we only consider a graph with neighborhood edges. So, one of the adjacent nodes of p, say, q must lie on Wp. Thus, we can write q 2 CðW p Þ, where CðW p Þ is the set of all intermediate nodes along Wp. Since (i) node q is a class A node and (ii) the shortest path (Wq) contains only such transit nodes which have higher probability to belong to object class, node q is an L1 node. h Theorem 2. If all the neighbor nodes of a particular node p are L2 or L3 nodes, then p cannot be an L1 node. Proof. As all the 8 neighbors q 2 Ne(p) are L2 or L3 nodes, then all shortest paths from u to q must contain at least one class B transit node or the node q itself must be a class B node. On the other hand, we know that any path from u to p must pass through at least one of the q nodes. So, we can never find any path between u and p which is devoid of a class B transit node. Therefore, p cannot satisfy the second condition of L1 node, i.e., along the shortest path from u

W 0p = (0, 0, 1) for the direct neighbor nodes of u (1, 1, 1) otherwise Lp = (0, 0, 1) for p = u (user specified node is an L1 node) (, , ) otherwise (labels not determined yet)

Step 2. Progress: Find the label Lp of a node p (wherep R T 1 ) which is closest to the node u using Table 1. Also, set Wp = Wip and T1 = T1 U {p} and save the total path. Step 3. Update: For all neighbor nodes q (whereq R T 1 ) of node p, set Wi+1q = min [W iq , (Wp + Lp)]. Repeat step 2 and step 3 until T1 = P. One can potentially choose any path between two equal paths i and j (niL 3 = njL 3 and niL 2 = njL 2 and niL 1 = njL 1 ). The connectivity term Cp can be mathematically defined as:

!,   X n   distðp; Ls Þ Cp ¼ X i exp  Ii  Ii1    i¼1 i¼1 n Y

ð13Þ

In Eq. (13), n is the number of nodes along the shortest path Wp and Xi is the state of the ith (i 2 CðX p Þ) transit node and its value dePn pends on its label (as shown in Table 2). The term i¼1 ðI i  Ii1 Þ

1473

A.K. Rudra et al. / Pattern Recognition Letters 34 (2013) 1470–1475 Table 2 Estimation of Xi in Eq. (13). Li

Xi

L1 L2 L3

1 Pr(Ip|O) 0

determines the signed intensity gradient along the path and dist(p,LS) is the distance between the node p and the closest L1 node (Ls) along the path Wp, in terms of the number of intermediate nodes. The time-complexity (Cormen et al., 2001) of the connectivity analysis algorithm, applied on the sparse graph G1(V1, E1) using binary heap, is O(E1 log V1). The time-complexity of the Edmonds–Karp enhanced Ford–Fulkerson algorithm, used to find a minimum cut is O(VE2). Since, E1  E andV 1  V, total time-complexity of the proposed method is O(E1 log V1) + O(VE2) = O(VE2). Fig. 1. Graph used for our algorithm: Red circles (2, 3, 5, 6) represent L1 nodes; Blue solid circle (7) represents L3 node; Black solid circles (1, 4, 8, 9) represent L2 nodes; node 6 has been marked as ‘u’; black line represents an N-link, blue lines represent a T-link, and, red lines (solid indicates strong connection and dotted indicates weak connection) represent C-links (4-neighbors shown for the purpose of clarity). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 1 Labeling of a node p. Wp

Pr(Ip|O)  Pr(Ip|B)

Lp

nL_3 = 0, nL_2 = 0, nL_1 > 0 nL_3 = 0, nL_2 = 0, nL_1 > 0 nL_3 = 0, nL_2 > 0, nL_1 P 0 nL_3 = 0, nL_2 > 0, nL_1 P 0 nL_3 > 0, nL_2 P 0, nL_1 P 0

>0 <0 >0 <0 –

L1 L2 L2 L3 L3

(0, 0, 1) (0, 1, 0) (0, 1, 0) (1, 0, 0) (1, 0, 0)

3. Experimental results We have so far experimented with 25 different T1-weighted MRI datasets with extremely poor contrast. In our protocol, gradient-echo T1 imaging was employed using a 1.5T MRI scanner (Signa Horizon LX Echo speed; General Electric Medical Systems, Milwaukee, WI, USA) with the use of a phased-array torso surface coil. Contrast agent gadoteric acid (Dotarem 0.5 mmol/mL; Guerbet, France) was bolus injected at a rate of 3–4 ml/s, at a dose of 0.2 ml/kgBW. Imaging parameters were: slice thickness = 5 mm, TR = 30–40 ms, TE = 2–3 ms, Flip angle 70°, F.O.V 38  38 cm2, and the matrix size 256  160. For each patient coronal single slice scan at the level of the renal hilum was taken before contrast. Approximately repeated 80 temporal frames of coronal scans were taken at 3-s intervals after contrast injection. For experiments in 2D, a single slice (of size 256  256 pixels) with extremely poor contrast is chosen from a 3D dataset for all

Fig. 2. Segmentation Results for two 2D Datasets. One dataset: (a) input with object seeds (in red) background seeds (in yellow), (b) output of Rudra et al. (2011) (D.C.: 68.92%), (c) output of Ali et al. (2007) (D.C.: 96.03%), (d) output of our method (D.C.: 98.78%). A Second Dataset: (e) input, (f) output of Rudra et al. (2011) (D.C.: 70.14%), (g) output of Ali et al. (2007) (D.C.: 97.18%), (h) output of our method (D.C.: 99.13%). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

1474

A.K. Rudra et al. / Pattern Recognition Letters 34 (2013) 1470–1475

Fig. 3. Variations of D.C. with parameters in the energy function. (a) Variations of D.C. vs. k for various fixed values of l. (b) Variations of D.C. vs. l for various fixed values of k. Table 3 Accuracy of different segmentation algorithms. Algorithm

Min. dice coeff.

Mean (dice coeff.) ± s.d. (dice coeff.)

Max. dice coeff.

Rudra et al. (2011) Ali et al. (2007) Our method

67.13 93.97 97.63

70.22 ± 1.83 96.02 ± 1.13 98.60 ± 0.52

73.22 97.84 99.57

the cases. The parameter k in is experimentally chosen to be 70 and the parameter l is set to 1. The value of the parameter r in Eq. (5), which can be treated as ‘‘camera noise,’’ is estimated as 3. In our proposed method, the only requirement for the user specified node u is that it should be within the object. We have compared the performance of our algorithm with two different graph cut-based segmentation methods described in Rudra et al. (2011) and Ali et al. (2007). In Rudra et al. (2011), graph cut was modified to segment brain MRI images suffering from poor contrast. So, we choose to compare the proposed method with that of Rudra et al. (2011) to emphasize the contribution of the connectivity induced energy term. The proposed connectivity-based approach is essentially a shape modeling technique built into the graph cut framework. Therefore, we decide to compare the method of kidney segmentation described in Ali et al. (2007) which employs graph cuts with shape priors. In Figs. 2 and 3, we show two sets of data with inputs (a), segmentation results of Rudra et al. (2011) (b), segmentation results of Ali et al. (2007) (c), and segmentation results of the proposed method (d). The contour of the kidney, obtained from the ground-truth, is shown in red and is superposed on each segmented output. The figures clearly demonstrate that the results in Fig. 1(d) and Fig. 2(d) are much closer in appearance to the respective ground-truths. Dice coefficient is used for quantitative comparison of the performance of these segmentation algorithms. Note that the D.C. is used as a standard measure for evaluating performance of an image segmentation algorithm. Importantly, this is an unbiased measure, i.e., it penalizes both over-segmentation and under-segmentation (Ge et al., 2006). For the dataset shown in Fig. 1(2), respective Dice coefficients using Rudra et al. (2011), Ali et al. (2007) and our method are 68.92% (70.14%), 96.03% (97.18%) and 98.78% (99.13%). Table 3 shows the minimum, mean ± s.d., and maximum Dice coefficient values of the above methods over the 25 datasets. In order to test the statistical significance of our improved accuracy over its competitors, we run the paired t-tests (Johnson and Bhattacharya, 2006) on the Dice coefficient data. The p-value obtained from the paired t-test on Dice coefficient data of the proposed method and that of Ali et al. (2007) is 1.47  1016. Similarly, the p-value resulted from the paired t-test on Dice coefficient data of the proposed method and that of Rudra et al. (2011) is 2.02  1033. Very low p-values in both the cases clearly suggest that the proposed method yields statistically significant performance improvement over both (Ali et al., 2007; Rudra et al., 2011). We have also calculated sensitivity and specificity values for our method. The average sensitivity for our method is 99.21% and the average specificity for our method is 99.88%. The proposed method is quite robust in nature both in terms of the sensitivity to the parameters k and l and sensitivity to the choice of the node ‘u’. We first show in Fig. 3 that our D.C. changes very little with variations in k and l in Eq. (2). We next illustrate using Fig. 4 the insensitivity of D.C. to the node ‘u’. The proposed method on average takes less than 5 s to execute on an

Fig. 4. Variations of D.C. with different positions of the node u.

A.K. Rudra et al. / Pattern Recognition Letters 34 (2013) 1470–1475

1475

exhibit very low contrast like ultrasound images in cardiology, breast cancer detection and study of growth of foetus (Noble and Boukerroi, 2006). References

Fig. 5. Segmentation Results for two 3D Datasets. One dataset: (a) ground-truth, (b) output of our method (D.C.: 98.27%). A second dataset: (c) ground-truth, (d) output of our method (D.C.: 98.49%).

Intel(R) Core(TM)2 Duo processor with a speed of 2.8 GHz. We have also experimented with five 3D datasets. On these five 3D datasets, we have obtained a (mean + s.d.) Dice Coefficient (D.C.) of (98 ± 0.43)% with 99.16% average sensitivity and 99.69% average specificity. In Fig. 5, we show the 3D segmentation results for two 3D datasets. 4. Conclusions and future work In this paper, we address a challenging biomedical problem of the segmentation of kidney from extremely low contrast MR images. A connectivity induced energy term is derived following the Dijkstra’s shortest path algorithm and is incorporated in the energy function described in Rudra et al. (2011). Minimization of this energy results in a high segmentation accuracy for the kidney images. The framework, discussed in this paper, provides a computationally inexpensive shape modeling technique without assuming any prior shape knowledge. Qualitative and quantitative results demonstrate that the proposed method has clearly outperformed the techniques described in Ali et al. (2007), Rudra et al. (2011). Since the proposed method does not assume any shape prior, we plan to apply this method to various other images which

Ali, A., Farag, A., El-Baz, A., 2007. Graph cuts framework for kidney segmentation with prior shape constraints. In: Ayache, N., Ourselin, S., Maeder, A. (Eds.), Proc. Intl. Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), 4791, pp. 384–392. Beatrice, C., Jean-Luc, C., Damien, M., Michel, C., Olivier, P., 2009. Functional semiautomated segmentation of renal DCE-MRI sequences using a growing neural gas algorithm. In: Zaher Ashraf, A. (Ed.), Recent Advances in Signal Processing. InTech, pp. 69–80. Boykov, Y., Funka-Lea, G., 2006. Graph cuts and efficient N–D image segmentation. Int. J. Comput. Vision 70, 109–131. Boykov, Y., Jolly, M.P., 2001. Interactive graph cuts for optimal boundary & region segmentation of objects in N–D images. In: Proc. IEEE Conf. on Computer Vision (ICCV), pp. 105–112. Chevaillier, B., Ponvianne, Y., Collette, J.L., Claudon, M., Pietquin, O., 2008. Functional semi-automated segmentation of renal DCE-MRI sequences. In: Proc. Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), pp. 525–528. Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C., 2001. Introduction to Algorithms. MIT Press. Ford, L.R., Fulkerson, D.R., 1962. Flows in Networks. Princeton University Press. Freedman, D., Zhang, T., 2005. Interactive graph cut based segmentation with shape priors. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 755–762. Freiman, M., Kronman, A., Esses, S.J., Joskowicz, L., Sosna, J., 2010. Non-parametric iterative model constraint graph min-cut for automatic kidney segmentation. In: Proc. Intl. Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), vol. 6363, pp. 73–80. Ge, F., Wang, S., Liu, T., 2006. Image-Segmentation evaluation from the perspective of salient object extraction. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 1146–1153. Hackjoon, S., Samuel, C., Cheng, T., Jin, H.W., Diana, K., Kyongtae, T.B., 2009. Semiautomated segmentation of kidney from high-resolution multidetector computed tomography images using graph-cuts technique. J. Comput. Assist. Tomogr. 33, 893–901. Johnson, R.A., Bhattacharya, G.K., 2006. Statistics: Principles and Methods. John Wiley & Sons. Khalifa, F., El-Baz, A., Gimel’farb, G., Abo El-Gahr, M., 2010. Non-invasive imagebased approach for early detection of acute renal rejection. In: Proc. Int. Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), vol. 6361, pp. 10–18. Lin, K.J., Huang, J.Y., Chen, Y.S., 2011. Fully automatic region of interest selection in glomerular filtration rate estimation from 99mTc-DTPA renogram. J. Digital Imaging 24 (6), 1010–1023. Malcolm, J., Rathi, Y., Tannenbaum, A., 2007. Graph cut segmentation with nonlinear shape priors. In: Proc. IEEE Conf. on Image Processing (ICIP), pp. 365–368. Noble, J.A., Boukerroi, D., 2006. Ultrasound image segmentation: a survey. IEEE Trans. Med. Imaging 25, 987–1010. Rigg, K.M., 1995. Renal transplantation: current status, complications and prevention. J. Antimicrob. Chemother 36, B51–B57. Rudra, A.K., Sen, M., Chowdhury, A.S., Elnakib, A., El-Baz, A., 2011. 3D graph cut with new edge weights for cerebral white matter segmentation. Pattern Recogn. Lett. 32, 941–947. Sabuncu, M.R., Thomas Yeo, B.T., Koen Van Leemput, Fischl Bruce, Golland Polina, 2009. Supervised nonparametric image parcellation. In: Proc. Intl. Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), vol. 5762, pp. 1075–1083. Slabaugh, G., Unal, G., 2005. Graph Cuts Segmentation Using an Elliptical Shape Prior. In: Proc. IEEE Conf. on Image Processing (ICIP), pp. 1222–1225. Veksler, O., 2008. Star shape prior for graph-cut image segmentation. In: Proc. European Conf. Computer Vision (ECCV), pp. 454–467. Vicente, S., Kolmogorov, V., Rother, C., 2008. Graph cut based image segmentation with connectivity priors. In: Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), Anchorage, AK, pp. 1–8. Yang, D., Ye, Q., Williams, M., Sun, Y., Hu, T.C.C., Williams, D.S., et al., 2001. USPIO enhanced dynamic MRI: evaluation of normal and transplanted rat kidneys. Magn. Reson. Med. 46, 1152–1163. Yuksel, S.E, El-Baz, A., Farag, A.A., 2006. Kidney segmentation framework for dynamic contrast enhanced magnetic resonance imaging. In: Proc. Int. Symposium on Mathematical Methods in Engineering (MME), pp. 1–10. Zöllner, F., Sance, R., Rogelj, P., Ledesma-Carbayo, M., Rørvik, J., Santos, A., Lundervold, A., 2009. Assessment of 3D DCE-MRI of the kidneys using nonrigid image registration and segmentation of voxel time courses. Comput. Med. Imaging Graph. 33 (171), 181.