Dynamic MR images are subject to relatively low signal-to-noise, nonuniform intensity distribution over the time series images, and geometric kidney deformations caused by gross patient motion, transmitted respiratory effects, and intrinsic and transmitted pulsatile effects. Therefore, accurate segmentation and registration of dynamic MR renal images is a challenge. These two basic steps are commanding the major attention in this research area for automated analysis of dynamic perfusion MRI. Particularly, kidney motion effects can be compensated for by specific use of global and local registration techniques. On the other hand, kidney segmentation techniques can be classified into three main categories: threshold-based, deformable boundary-based, and probabilistic or energy minimization-based methods. Below we review the related work on kidney segmentation and registration techniques addressing the above-mentioned challenges.

Threshold-based techniques segment the kidney and its internal structures (i.e., cortex and medulla) by analyzing an empirical probability distribution, or histogram of pixel intensities in a region-of-interest (ROI). Earlier computerized renal image analysis (e.g., [51–54]) was usually carried out either manually or semiautomatically. Typically, the user defines an ROI in one image and for the rest of the images, image edges were detected and the model curve was matched to these edges. However, the manual ROI placements are based on the users’ knowledge of anatomy and thus are subject to inter- and intra-observer variability. Also, valuable information, being inherent in the DCE-MRI signal intensity time series (sequences), is not used. Additionally, these approaches are very slow, even though semiautomated techniques (e.g., [51, 54]) do reduce the processing time. Giele et al. [55] introduced an approach for the segmentation and registration and of the kidney on DCE-MRI. First, the kidney contour is drawn manually by the user in a single high-contrast image. Then, the phase difference movement detection (PDMD) method is employed to correct kidney displacements. Their method demonstrated better performance than direct image intensity matching and cross-correlation. However, when compared with the radiologist results, the PDMD method accuracy was about 68 % and a manual mask to register the time frames was still required. Additionally, only translational motion was handled, while rotational motion was not mentioned, although the existence of the latter has been discussed in a number of studies [51, 56, 57]. De Priester et al. [54] subtracted the average of pre-contrasted images (10 frames) from the average of early enhanced images (30 frames) and thresholded the resulting difference image to obtain a kidney mask. Objects smaller than a certain size (700 pixels) were removed, and the remaining kidney object was closed using morphological erosion and manual processing. This approach was further expanded by Giele [58] by applying an erosion filter to the mask image in order to obtain a contour at a second subtraction stage. Koh et al.[59] segmented kidneys with the morphological 3D H-maxima transform. Rectangular masks and edge information are used to exclude training data or prior knowledge. Simple thresholding is too inaccurate to segment human organs in DCE-MRI, because these specific regions have similar gray level (intensity) distributions.

Evolving deformable boundary methods have been explored as a more accurate means of kidney segmentation. A series of studies on both rats and human subjects [57, 60–65] has been conducted for the registration and segmentation of kidneys from DCE-MRI. A multi-step segmentation and registration in the study on humans by Sun et al. [57, 62] initially corrects the large-scale motion by using an image gradient-based similarity rigid registration (only translational). Once roughly aligned, a high-contrast image is subtracted from a pre-contrast image and a level set approach was used to extract the kidney border from the difference image. Then, the segmented kidney contour is propagated over the other frames to search for the rigid (rotation and translation) registration parameters. For rat studies, Sun et al. [60, 61, 63] used a variational level set approach to find the kidney borders. Their framework integrated a subpixel motion model and temporal smoothness constraints. For segmenting the cortex and medulla, the level set approach by Chan and Vese [66] was used. The deformable model-based segmentation has been improved by hybrid edge- and region-based models in a number of studies [67, 68]. Some works have focused on segmenting multiple objects with multiphase level set methods [69, 70]. Abdelmunim et al. [71] incorporated both image and shape prior information into a variational level set framework for kidney segmentation. However, their model did not adequately account for spatial dependencies between the pixels and therefore are quite sensitive to imperfect kidney contours and image noise. Yuksel et al. [72, 73] proposed a parametric deformable model approach for the segmentation of the kidney where the deformable contour evolution was constrained using two density functions. The first describes the kidney shape prior and is constructed using the average signed distance maps of the training samples. The second functional describes the pixel-wise image intensity distribution of the kidney and its background, estimated using an adaptive linear combinations of discrete Gaussians (LCDG) [74–81]. El-Baz et al. [82–84] proposed a parametric deformable model-based approach for the segmentation of the kidney using shape and visual appearance priors. The shape model is constructed from a linear combination of vectors of distances between the training boundaries and their common centroid. The appearance prior is modeled with a spatially homogeneous second-order Markov-Gibbs random field (MGRF) of gray levels with analytically estimated pairwise potentials. The current appearance model is described with the LCDG model [74–81]. Khalifa et al. [29, 85] proposed an automated level set-based framework for the segmentation of kidney from dynamic MRI. They proposed a stochastic force that accounts for a shape prior and features of image intensity and pairwise MGRF spatial interactions [75, 86–88]. These features are integrated into a joint MGRF image model of the kidney and its background to constrain the evolution of the deformable contour [89–92]. They employed a two-stage registration methodology using first an affine transformation to account for the global motion, followed by a partial differential equation (PDE)-based approach for local motion correction [90, 92]. The segmentation approach in [29, 89] was later extended to deal with 3D data in [93, 94]. Gloger et al. [95] presented a level set-based approach using the shape prior information and Bayesian statistical concepts for generating the shape probability maps. However, the shape prior model in [29, 89, 93–95] did not impose temporal constraints on kidney segmentation.

The graph cut-based segmentation algorithm by Boykov et al. [96, 97] minimizes the energy of a temporal MGRF model of intensity curves. Each voxel is described with a vector of intensity values over time. Initially, several seed points are placed on the objects and on the background to give user-defined constraints as well as expert samples of intensity curves. These samples are used to compute a two-dimensional histogram further acting as a data penalty function in minimizing the energy. Although the results looked promising, manual interaction was still required. Rusinek et al. [98] proposed a graph cut-based segmentation formwork to assess cortical and medullary functional parameters. Their method employed a rigid registration step to account for the kidney displacements and the approach has been tested on simulated and in-vivo data. Ali et al. [99] used the graph cut-based minimization of an energy functional combining a shape constraint with boundary properties. The constraint was built using a Poisson probability distribution and distance maps. Chevaillier et al. [100, 101] proposed a semiautomated method to segment internal structures (i.e., cortex, medulla, and pelvis) from DCE-MRI time series by using k-means-based partitioning to classify pixels according to contrast evolution using vector quantization algorithm. However, it was only tested on eight data sets for normal kidneys, and user interaction was still required. A similar segmentation by Song et al. [102] has only been tested on two MRI data sets, with simulated rotation and translation rigid motion, for one normal and one abnormal kidney. An automated framework proposed by Zöllner et al. [103] assesses renal function by deriving voxel-based functional information from DCE-MRI, the nonrigid image registration compensating for the motion and deformation of the kidney during DCE-MRI acquisitions. The k-means clustering method [104] was used for extracting functional information about different regions of the kidney according to their dynamic contrast enhancement patterns. An automated wavelet-based k-means clustering framework for segmenting the kidneys was proposed by Li et al. [105]. The images were co-aligned using B-splines registration and cross-correlation (CC) cost function and their framework was tested on seven subjects (four volunteers and three patients). Yang et al. [106] proposed a framework for the classification of kidney tissue using fuzzy c-mean clustering. In order to reduce the motion artifacts, their framework employed a nonrigid registration step using the demons algorithm [107] and the squared pixel distance as a similarity metric and the squared gradient of the transformation field as the smoothness regularization term. To summarize, the reviewed methodologies for kidney segmentation and registration are presented in Table 2.