Abstract
Vessel segmentation algorithms are critical components of circulatory blood vessel analysis systems. We present a survey of vessel extraction techniques and algorithms, putting the various approaches and techniques in perspective by means of a classification of the existing research. While we target mainly the extraction of blood vessels, neurosvascular structure in particular, we also review some of the segmentation methods for the tubular objects that show similar characteristics to vessels. We divide vessel segmentation algorithms and techniques into six main categories: (1) pattern recognition techniques, (2) model-based approaches, (3) tracking-based approaches, (4) artificial intelligence-based approaches, (5) neural network-based approaches, and (6) miscellaneous tube-like object detection approaches. Some of these categories are further divided into sub-categories. A table compares the papers against such criteria as dimensionality, input type, pre-processing, user interaction, and result type.
1 Introduction
Blood vessel delineation on medical images is essential to solving several practical applications such as vessel diagnosis (e.g. stenosis or malformations) and registering patient images obtained at different times. Segmentation methods vary depending on the imaging modality, application domain, method being automatic or semi-automatic, and other specific factors. No single segmentation method can extract vasculature from every medical image modality. While some methods employ pure intensity-based pattern recognition techniques such as thresholding followed by connected component analysis (CCA) [1], some other methods apply explicit vessel models to extract the vessel contours[2]. Depending on the image quality and the general image artifacts such as noise, some segmentation methods may require image preprocessing prior to the segmentation algorithm [3].Some other methods apply post-processing to overcome the problems arising from over segmentation. We divide vessel segmentation algorithms and techniques into six main categories: (1) pattern recognition techniques, (2) model-based approaches, (3) tracking-based approaches, (4) artificial intelligence-based approaches, (5) neural network-based approaches, and (6) miscellaneous tube-like object detection approaches. Pattern recognition techniques are further divided into seven categories: (1)multi-scale approaches, (2) skeleton-based approaches, (3)region growing approaches, (4) ridge-based approaches, (5)differential geometry-based approaches, (6) matching filters approaches, and (7) mathematical morphology schemes.
Model-based approaches are also further divided into four categories: (1) deformable models, (2) parametric models, (3) template matching approaches, and (4) generalized cylinders approaches. Although we divide segmentation methods in different categories, sometimes multiple techniques are used together to solve different segmentation problems. We, therefore, cross-listed the methods that fall into multiple segmentation category in the comparison table given at the end of the paper in Table 1.We survey current vessel segmentation methods, covering both early and recent literaturerelated to vessel segmentation algorithms and techniques. We introduce each segmentation method category and briefly summarize papers by category. Due to the space limitation, we present only a subset of the papers in each group. A more comprehensive review, set of comparison tables, and complete list of references are posted on the web [4]. We provide a table that compares the papers against such criteria as dimensionality,input type, pre-processing, and user interaction.
2 Pattern Recognition Techniques
Pattern recognition techniques deal with the automatic detection or classification of objects or features. For vessel extraction, pattern recognition techniques are concerned with the automatic detection of vessel structures and features. We divide pattern recognition techniques into seven
categories: (1) multi-scale approaches, (2) skeleton-based (centerline) approaches, (3) region growing approaches, (4) ridge-based approaches, (5) differential Geometry-based approaches, (6) matching filters approaches, and (7) mathe-matical morphology schemes.
2.1 Multi-scale Approaches (MSA)
Multi-scale approaches perform segmentation at varying image resolutions. The main advantage of this technique is increased processing speed. Major structures (large vessels in our application domain) are extracted from low resolution images while fine structures are extracted at high resolution. Another advantage is increased robustness. After segmenting the strong structures at the low resolution, weak structures, such as branches, in the neighborhood of the strong structures can be segmented at higher resolution.
Sarwal and Dhawan [5] reconstruct 3D coronary arteries from three views by matching branch points in each view.Their method is based on simplex method-based linear programming and relaxation-based consistent labeling. To improve the robustness of the matcher, matching process is performed at three different resolutions. The stronger vessel tree branches are extracted at high resolution while the weaker branches are extracted at lower scale. The extracted vessel tree is used for 3D reconstruction.Chwialkowski et al [6] employ multiresolution analysis based on wavelet transform. Their work aims at automated qualitative analysis of arterial flow using velocity-sensitive,phase contrast MR images. The segmentation process is applied to the magnitude image and the velocity information from the phase difference image is integrated on the resulting vessel area to get the blood flow measurement. Vessel boundaries are localized by employing a multivariate scoring criterion to minimize the effect of imaging artifacts such as partial volume averaging and flow turbulence.
2.2 Skeleton-based Approaches (SBA)
Skeleton-based methods extract blood vessel centerlines.The vessel tree is created by connecting these centerlines.Different approaches are used to extract the centerline structure. Variously these methods apply thresholding and then object connectivity, thresholding followed by thinning procedure, and extraction based on graph description. The resulting centerline structure is used for 3D reconstruction.
Tozaki et al [7] extract bronchus and blood vessels from thin slice CT images of the lung for 3D visualization and analysis. First, a threshold is used to segment the images.Then, blood vessels and bronchus are differentiated by using their anatomical character (e.g. the bronchus contain air). Finally, a 3D thinning algorithm is applied to extract the vessel centerlines. The resulting centerline structure is used to analyze and classify the blood vessels. Their work
helps in early detection of tumors of lung cancer patients.
Kawata et al [8] analyze blood vessel structures and detect blood vessel diseases from cone-beam CT images. X-ray digital angiograms are collected using rotational angiography. 3D image reconstruction is performed by a short scan cone-beam filtered backprojection algorithm based on the short injection time of the contrast medium. First, a graph description procedure extracts the curvilinear centerline structures of the vessel tree using thresholding, elimination of the small connected components, and 3D fusion processes. Then, a 3D surface representation procedure extracts the characteristics of convex and concave shapes on blood vessel surface. The algorithm is run on a set of patient images with abdominal blood vessels, with two aneurysms and one stenosis, and the results are shown.
2.3 Ridge-based Approaches (RBA)
Ridge-based methods treat grayscale images as 3D elevation maps in which intensity ridges approximate the skeleton of the tubular objects [9]. Ridge points are local peaks in the direction of maximal surface gradient, and can be obtained by tracing the intensity map from an arbitrary point, along the steepest ascent direction. Ridges are invariant to affine transformations and can be detected in different image modalities. These properties are exploited in medical image registration [10, 11]. Since RBA detect skeleton of tubular objects, it can be thought of as a specialized SBA.
Bullitt and Aylward [12] describe their method of defining vessel trees from 3D image volume. The segmentation stage starts with a manually-selected seed point for each vessel. The system extracts an intensity ridge map which represents the vessel’s medial axis. Vessel width at each ridge point is also calculated using a scale-based approach.The vessel tree is represented by a graph where each vessel keeps information about its relationship to other vessels.The main application of this work is in the registration of vasculature images obtained from the same patient at different times. Such registration allows the observation of changes in pathology over time.
Aylward et al [10] approximate the medial axes of tubular objects such as vessels in an angiogram as directed ‘intensity ridges’. They apply the cores method [13] which has been proven to be invariant to a wide range of noise and object disturbances. Ridges are tracked by estimating the local vessel directions. First, image intensity is mapped to height to create intensity height surface. Second, from a user-supplied starting point ,an initial ridge point is found using a conjugate directions search with respect to the Hessian matrix. Third, the ridge is tracked. Finally, the local widths of the segmented object is estimated using points on the ridges. The authors show results of a vascular tree ex-tracted from a MR angiogram. This required a fair amount of user intervention (105 mouse clicks in all). Figure 1 shows the extracted vascular tree.
2.4 Region Growing Approaches
Starting from some seed point, region growing (RG) techniques segment images by incrementally recruiting pixels to a region based on some predefined criteria. Two important segmentation criteria are value similarity and spatial proximity. It is assumed that pixels that are close to each other
and have similar intensity values are likely to belong to the same object. The main disadvantage of region growing approach is that it often requires user-supplied seed points. Due to the variations in image intensities and noise, RG can result in holes and over segmentation. Thus, it requires post-processing of the segmentation result.
O’Brien and Ezquerra [14] automatically segment coronary vessels in angiograms based on temporal, spatial, and structural constraints. The algorithm starts with a low pass filtering applied to the image as preprocessing. Then, initial segmentation starts with a user-supplied seed point.The system starts a RG process to extract the initial approximation to the vessel structure. After that, the centerlines are extracted by employing a balloon test . Next, undetected vessel segments are located by a spatial expansion algorithm. At this stage, images are divided into two categories: vessel areas and non-vessel areas. However, there is no spatial or temporal connectivity information exists in the detected sub-regions. This information is extracted by applying an acceptance and rejection test using graph theory. Due to centerlines extraction, this work can also be classified as a SBA listed in section 2.2.
Higgins et al [15] describe their automatic arterial tree extraction algorithm from 3D coronary angiograms. These angiograms are obtained from high-resolution X-ray CT scanner known as 3D Dynamic Spatial Reconstructor (DSR). Their algorithm is a combination of a 3D filter, a CCA, a thresholding process, and seeded RG algorithm. The strength of the algorithm is reported as the results being reproducible, requiring less user time, and working in 3D. Due to the skeleton detection process performed, this work can also be classified as a SBA listed in section 2.2.
2.5 Differential Geometry-based Approaches
Differential geometry (DG) based methods treat images as hypersurfaces and extracts features using the curvature and the crest lines of the surface. The crest points of the hyper-surface correspond to the center lines of the vessel structure. The 2D and 3D images are treated similarly, being modelled as 3D and 4D hypersurfaces respectively. In DG a 3D surface can be described by two principal curvatures and by their corresponding orthogonal directions, called principal directions. These features are also invariant under affine transformations and therefore well-suited to medical image registration. The principal curvatures correspond to the eigenvalues of the Weingarten matrix and the principal directions are the eigenvectors. Crest points, which are the intrinsic properties of the surfaces, are the local maxima of Figure 2. a. DSA image of the cerebral vessels, and b. Vessel detection using four values of the scale (Reproduced from [17]) the maximum curvature on the hypersurface. Center-lines are obtained by linking the crest-points.
Prinet et al [16] propose a multidimensional vessel extraction method that treats images as parametric surfaces and extracts features of the images using surface curvature and the crest lines. When linked together, the crest points form the center lines of the vessels. Results of the algorithm applied to angiograms, 2D Digital Subtraction Angiography (DSA), Magnetic Resonance Angiography (MRA), and 3D synthetic data are reported.
Armande et al [17] extract thin nets using a MSA that exploits the DG properties of the image surface. They characterize thin nets as crest lines of the image surface. To overcome the problem faced in extraction of the thin nets having different widths, they employ a MSA. Their method consists of three main stages: (1) They detect the extrema of the maximum curvature for all scales; (2) They remove false responses, using the gradient zero-crossings; and (3) They select those points verified by medium scale expression as crest points. Figure 2 shows a DSA image and the extracted vessel network using four different scales. This work can also be classified as a MSA listed in section 2.1.
2.6 Matching Filters Approaches
Matching filters (MF) approach convolves the image with multiple matched filters for the extraction of objects of interest. In extracting vessel contours, designing different filters to detect the vessels with different orientation and size plays a crucial role. The convolution kernel size affects the computational load. MF are usually followed with some other image processing operations like thresholding and CCA to get the final vessel contours. CCA is preceded by a thinning process to detect vessel centerlines.
Sato et al [3] introduce a 3D multi-scale line enhancement filter to segment curvilinear structures in medical images. The filter is based on the directional second derivatives of smoothed images using Gaussian kernel in multi scales with adaptive orientation selection using the Hessian matrix. They demonstrate the result of their method applied to segment brain vessels from MRI/MRA and bronchi from a chest CT, and liver vessels from an abdominal CT.
Hoover et al [18] combine local and region-based properties to segment blood vessels in retinal images. The method examines the matched filter response (MFR) [19], using a probing technique. The technique classifies pixels in an area of the MFR as vessels and non-vessels by iteratively decreasing the threshold. At each iteration, the probe examines the region-based attributes of the pixels in the tested area and segments the pixels classified as vessels. Pixels that are not classified as vessel from probes are recycled for further probing. A unique feature of this method is that each pixel is classified using local and region-based properties.
2.7 Mathematical Morphology Schemes
Morphology relates to the study of object forms or shapes. Morphological operators (MO) apply structuring elements (SE) to images, and are typically applied to binary images but can be extended to the gray-level images. Dilation and erosion are the two main MO. Dilation expands objects by a SE, filling holes, and connecting disjoint regions. Erosion shrinks objects by a SE. Closing, dilation followed by erosion, and opening, erosion followed by dilation, are two other operations. Two algorithms used in medical image segmentation and related to mathematical morphology are top-hat and watershed transformations [20].
Eiho and Qian [21] propose a method based on pure MO to detect coronary artery tree in cine-angiograms. The steps of the method are: (1) A top-hat operator is applied to enhance the shape of the vessels; (2) Morphological erosion followed by half-thresholding operations are applied to remove the areas that are not the coronary artery; (3) Starting from a user-supplied point on the tree, the system extracts whole tree using neighbor checking according to the average gray levels; (4) The extracted artery tree is skeletonized by the thinning operation. (5) The edges are extracted by applying watershed transformation on the binary image obtained from a dilation operation on the binary skeleton. This work can also be classified as a SBA listed in section 2.2.
Thackray and Nelson [22] describe an approach which extracts vascular segments using a set of 8 MO, each of which represents an oriented vessel segment. The system also applies an adaptive thresholding scheme to extract the vascular segments from the intensity image. The system was used to extract vessel segments in a capillary angiogram of mice, and does not extract the vascular interconnection structure. The range of vessel widths the system handles appears limited by the setting of the 8 MO.
3 Model-Based (MB) Approaches
MB approaches apply explicit vessel models to extract the vasculature. We divide MB approaches into four categories: (1) Deformable models, (2) Parametric models, (3) Template matching, and (4) Generalized cylinders.
3.1 Deformable Models (DM)
We divide DM into two categories: parametric DM and geometric DM. These categories are discusses in detail in the next sections. A survey on DM in medical image analysis is published by McInerney et al [23].
3.1.1 Parametric DM
DM are MB techniques find object contours using parametric curves that deform under the influence of internal and external forces. First introduced by Kass et al [24], active contour models or snakes are a special case of a more general technique of matching a DM by means of energy minimization.Physically, a snake is a set of connected control points, called snaxels . Each snaxel has an associated energy that either rises or falls depending upon the forces that act on it. Internal forces serve to impose smoothness constraints on the contour while external forces pull the snake towards the desired image features like lines and edges. The smoothness constraint imposed by elasticity energy makes the DM robust to the noise. The main disadvantage is that usually it requires user interaction to initialize the snake. It also requires initial parameters given by the user.
Luo et al [25] design a model that overcomes the problems associated with traditional snakes, such as contour initialization, internal parameter setting, and the limitations in the capture range of the external energy (EE). The model has new internal energy (IE) and new EE that are treated equally. The IE maintains smoothness without any shrinking side effects on the contour. This is accomplished by computing “just enough” smooth force to overcome the image force. The EE combines both edge and region information. This reduces the effects of contour initialization. The model was tested on both synthetic and real gray-level images and reported encouraging results.
Toledo et al [26] combine a probabilistic principal component analysis technique (PPCAT) with a statistical snake (SS) technique to track non-rigid elongated structures. PP-CAT is used to construct statistical image feature descriptions while snakes are used for global segmentation and to track objects. The SS learns and tracks image features using statistical learning techniques. A likelihood map, used by SS, is created from a training set of object profiles using the PPCAT. Each point in the map is assigned a probability measure to belong to the learned feature category. The likelihood map is extended, by applying an extended local coherence detection to the coherent direction field, to give priority to parallel coherent structures. The likelihood map is used to define a probabilistic potential field of the snake. The SS deforms itself to maximize the overall probability of detecting learned image features.
3.1.2 Geometric DM and Front Propagation Methods
Caselles et al [27] and Malladi et al [28] use propagating interfaces under a curvature dependent speed function to model anatomical shapes. They use the Level Set Method (LSM) approach developed by Osher and Sethian [29] and adapt it to shape recognition. The main idea behind the LSM is to represent propagating curves as the zero level set of a higher dimensional function which is given in the Eulerian coordinate system. Hence, a moving front is captured implicitly by the level set function (LSF). The advantages of this approach are: (1) It can handle complex interfaces which develop sharp corners and change its topology during the development; (2) Intrinsic properties of the propagating front such as the curvature and normal to the curve can be easily extracted from the LSF; (3) Since the LSF is given in the Eulerian coordinate system, discrete grids can be used together with finite differences methods to obtain a numerical approximation to the solution; and (4) It is easily extendable to higher dimensions.
Sethian developed a computationally efficient Fast Marching Method (FMM) [30], which uses a wave propagation (WP) approach for specialized front problems. FMMs are used in the problems where the front advances monotonically with a speed that does not change its sign.
Quek and Kirbas [31] develop a WP and traceback mechanism to extract vasculature from angiogram images. Using a dual-sigmoidal filter, the system labels each pixel in an angiogram with the likelihood that it is within a vessel. Representing the reciprocal of this likelihood image as an array of refractive indices, a digital wave is propagated through the image from the base of the vascular tree. This wave ‘washes’ over the vasculature, ignoring local noise perturbations. The extraction of the vasculature becomes that of tracing the wave along the local normals to the waveform. While the approach is inherently SIMD, they present an efficient sequential algorithm for the WP, and discuss the traceback algorithm. An example of WP is shown in Figure 3.
3.2 Parametric Models (PM)
PM approaches define objects of interest parametrically. In tubular object segmentation, objects are described as a set of overlapping ellipsoids. Some applications use a circular vessel model [32]. The parameters of the model used are estimated from the image. While an elliptic PM can approximate healthy vessels and stenoses, it fails to approximate pathological irregular shapes and vessel bifurcations.
Chan et al [32] utilize diameter information contained within the intensity profile amplitude(IPA) to estimate diameters of narrow vessels in X-ray cine-angiograms. A unique feature of the IPA is that it is sensitive to changes in small vessel diameters in case of noise and blur. The method has two steps: (1) Estimation of the imaging model parameters directly from the images and estimation of the diameters from these parameters. This step has three components to achieve imaging model parameters: a circular vessel model, a nonlinear imaging model, and a parameter estimation. (2) Application of a maximum likelihood (ML) estimation technique with amplitude information incorporated. It is reported that the model successfully estimates the diameters in the range of 0.4 mm to 4.0 mm.
Krissian et al [33] develop a multiscale model to extract and reconstruct 3D vessels. It consists of three main steps: (1) Multiscale responses from discrete set of scales is computed; (2)Local extrema in multiscale response is extracted; and (3) Skeleton of the local extrema is created and the result is visualized. A cylindrical vessel model is utilized in the first step to interpret the eigenvalues of the Hessian matrix and to choose a good normalization parameter. The initial tests gives promising results, with some local problems at vessel junctions and tangent vessels.
3.3 Template Matching
Template matching tries to recognize a structure model (template) in an image. The method uses the template as a context ,which is apriori model. Thus, it is a contextual method and a top-down approach. In arterial extraction applications, the arterial tree template is usually represented in the form of a series of nodes connected in segments. This template is then deformed to fit the structures in the scene optimally. Stochastic deformation process described by a
Hidden Markov Model (HMM) is a method to achieve template deformation [34]. Dynamic programming is an effective method employed in recognition process.
Summers and Bhalerao [35] implement a multi-resolution technique based on octree representation for the segmentation of MRA. The image data is first expanded in an octree representation using averaging on the combined set of velocity component images. Image blocks, that pass the confidence test for the occupancy probability and coherence test for adjacency, form the segmented tree. The system estimates features like flow direction, vessel axis,
diameter, and velocity from the segmented blocks using the local pressure gradient. The model is tested in extraction of vessels from MRA images and in calculation of pressure gradients in a model stenosis.
Van der Weide et al [36] localize paramagnetic markers to localize intravascular devices in MR images. The aim is to support the MR-guided vascular interventions. The method has two main steps: (1) Marker candidates, which are local minima (“blobs”), in the image are detected using both Laplacian image and winding number image. Winding number image is used to topologically classify different singular points such as local minima and local maxima points; (2) The intravascular device is identified by a matching process of the detected marker pattern to the known template of the device. The results of an animal experiment is discussed and 95 % of success rate is stated in phantom experiments.
3.4 Generalized Cylinders Model
Generalized cylinders (GC) are used to represent cylindrical objects. Technically GC are PM but we discuss them separately because there is a significant amount of work on this model and because of its prominence in the literature. Binford [37] introduced the use of GC in vision applications. GC consists of a space curve, or axis, and a cross-section function defined on that axis. Cross-section function is usually an ellipse. Tubular objects are defined by a
cross-sectional element that is swept along the axis of the tube (spine) using some sweep rules. The spine is represented by a spline and the cross-section function is represented ellipse. Another method to represent cylinders is to use Frenet-Serret formulation as the basis of GC [38]. However, Frenet-Serret formulation model and tube model described earlier suffer some serious drawbacks, such as discontinuities and non-intuitive twisting behavior.
O’Donnel et al [39] use a form of GC to recover cylindrical structures from medical images. A GC is a volume created by cross-section swept along a path, the spine. The spine is represented by a 3D cubic B-spline and the cross-section swept is always in the plane orthogonal to the spine to form the cylinder. The strength of their model comes from additional finite element (FEM) mesh-like component lying on top of their model to address the fine detail in complex structures. Figure 4 shows a result of their approach.
Sato et al [40] propose a new semi-automated method based on multi-scale Hessian-based technique to determine the position, orientation, and diameter of stenoses in coronary angiograms. The Hessian matrix, H, describes the second-order structure of local intensity variations around each point in the image. The method consists of five stages: (1) Two images in which stenosis can be seen are selected; (2) corresponding points in two images are manually selected to find translational parameters; (3) 2D positions and orientations of the stenosis in two images are estimated; (4) 3D position and orientation of the stenosis are calculated based on the principle of binocular stereo; (5) The vessel of interest with stenosis and any peripheral vessels which may be overlap the stenosis are specified manually. The method utilizes scale-dependency to formulate the diameter estimation to reduce user interaction.
4 Tracking-Based Approaches
Tracking-based approaches apply local operators on a focus known to be a vessel and track it. On the other hand, pattern recognition approaches apply local operators to the whole image. Vessel tracking (VT) approaches, starting from an initial point, detect vessel centerlines or boundaries by analyzing the pixels orthogonal to the tracking direction. Different methods are employed in determining vessel contours or centerlines. Edge detection operation followed by sequential tracing by incorporating connectivity information is a straightforward approach. Aylward et al [10] utilize intensity ridges to approximate the medial axes of tubular objects such as vessels. Some applications achieve sequential contour tracing by incorporating the features, such as vessel central point and search direction detected from previous step into the next step [41]. Fuzzy clustering is another approach to identify vessel segments. It uses linguistic descriptions like “vessel” and “nonvessel” to track vessels in retinal angiogram images. After the initial segmentation, a fuzzy tracking algorithm is applied to each candidate vessel region. Some methods utilize a model in the tracking process and incrementally segment the vessels. A more sophisticated approach on VT is the use of graph representation [43]. The segmentation process is, then, reduced to finding the optimum path in a graph representation of the image. One disadvantage of the VT approaches is that they are not fully automatic and require user intervention for selecting starting and end points.
Tolias and Panas [42] develop a fuzzy C-means (FCM) clustering algorithm that uses linguistic descriptions like “vessel” and “nonvessel” to track fundus vessels in retinal angiogram images. Their algorithm uses only (fuzzy) image intensity information and makes no assumptions for the shape of the vessels. First, optic nerve in fundus images is detected and used as the starting point. Next, the bounding circle of the optic nerve is found. Then, the points in the bounding circle are segmented as “vessel” and “nonvessel” using a FCM. Finally, a fuzzy vessel tracking algorithm is applied to each candidate vessel. The algorithm does not utilize any edge information to locate the vessels and this reduces the effects of noise in the tracking procedure.
Lecornu et al [43] extract vessel contours in angiogram images by tracking two edges simultaneously by means of graph theory. A blood vessel model is incorporated and some blood vessel properties such as the position and size of section and the curvature of the segment are used in the formal structure model. The detection process employs a heuristic search method which searches for the best edge in an image. The best edge is found as the optimum path in a graph representation of the image. They improve the algorithm by improving node concept by considering two opposite edges together to represent the vessel segments.
5 Artificial Intelligence-Based Approaches
Artificial Intelligence-based approaches (AIBA) utilize knowledge to guide the segmentation process and to delineate vessel structures. Different types of knowledge are employed in different systems from various sources. One knowledge source is the properties of the image acquisition technique, such as cineangiography, DSA, computed tomography (CT), MRI, and MRA. Some applications utilize a general blood vessel model as a knowledge source. Smets et al [45] encode general knowledge about appearance of blood vessels in the form of 11 rules (e.g. that vessels have high intensity center lines, comprise high intensity regions bordered by parallel edges etc.). Stansfield [46] applies a domain-dependent knowledge of anatomy to interpret cardiac angiograms in the high-level stages. According to Stansfield, “Anatomical knowledge is embodied within the system in the form of spatial relations between objects and the expected characteristics of the objects themselves.” Knowledge-based systems exploit apriori knowledge of the anatomical structure. These systems employ some low-level image processing algorithms, such as thresholding, thinning, and linking, while guiding the segmentation process using high-level knowledge. AIBA perform well in terms of accuracy, but the computational complexity is much larger than some other methods.
Rost et al [47] describe a knowledge-based system, called SOLUTION (Solution for a Learning Configuration System for Image Processing), designed to automatically adopt low-level image processing algorithms to the needs of the application. It aims to overcome the problem of extensive change requirement in the existing system to perform in a different environment. The system accepts task descriptions in high-level natural spoken terms and configures
the appropriate image processing operators by using expert knowledge formulated explicitly by rules. In the present implementation, extraction process is limited to contours.
6 Neural Network-Based Approaches
Neural networks(NN) are used to simulate biological learning and widely used in pattern recognition. Neural nets are basically a classification approach. The network is a collection of elementary processor (nodes). Each node takes a number of inputs, performs elementary computations, and generates a single output. Each node is assigned a weight and the output is a function of weighted sum of the inputs. These weights are learned through training and then used in the recognition. Back-propagation algorithm is a widely used learning algorithm. One of the advantages that make neural networks attractive in medical image segmentation is their abilityto use nonlinear classification boundaries obtained during the training of the network. Another attractive feature of the neural nets is the ability to learn. With the selection of a good training set which includes all possible features or objects, the network can learn the classification boundaries in its feature space. One of the disadvantages of NN is that they need to train every time a new feature is introduced the network. Another limitation is that it is difficult to debug the performance of the network.
Nekovei and Sun [48] describe their back-propagation network for the detection of blood vessels in X-ray angiog-raphy. This system does not extract the vascular structure. Its purpose is to label the pixels as vessel or non-vessel. The system applies the neural network directly to the angiogram pixels without prior feature detection. Since angiograms are typically very large, the network is applied to a small subwindow which slides across the angiogram. The pixels of this subwindow are directly fed as input to the network. Pre-labeled angiograms are used as the training set to set the network’s weights. A modified version the common delta-rule is to obtain these weights. The algorithm is compared with two other algorithms,bayesian maximum likelihood algorithm and iterative ternary classification algorithm .
Shiffman et al [49] combine an automated neural network-based segmentation approach with manual editing to extract sections from CT angiography (CTA) image volumes. They aim to facilitate the visualization of vasculature by editing the target sections in the volume prior to 3D reconstruction. In the first step, entire image sequence is segmented automatically producing a set of labelled image sections. The next step requires the user to view the resulting images and edit one or more sections by pointing and clicking within each region of interest. In the final step, user edited segments and the remaining section are connected to extract the final image segments based on label identity. Automated segmentation is achieved in two steps; a multilevel thresholding and then smoothing the resulting fuzzy regions. Two clustering methods, K-means clustering algorithm and a NN-based algorithmbased on Kohonen’s self-organizing feature maps, has been implemented and the results are compared in this work.
7 Miscellaneous Tube-Like Object Detection
This class of research approaches deals with the extraction of tubular structures from images. This is actually a ”mis-cellaneous” class of approaches that may be applicable to vascular extraction in that vessels are tubular entities, but these approaches were not designed for vessel extraction. Mayer et al [50] develop a model for the extraction of roads from aerial images. Their model has three basic components. First, multi-scale modeling is used to combine fine scale detailed information, such as the road markings, with coarse scale abstract information, such as the road network. Second, context information in the form of relations to other objects such as buildings and trees is exploited to extend the model. Using context information facilitates the extraction process to focus on the target objects. Third, ribbon-snakes are used to extract roads in fine scales. Using ribbon-snakes is reported to help the extraction of the roads occluded by shadows cast by building and trees in the image.
Huang and Stockman [51] describe a system that uses GC Generalized cylinders to extract tubular structures in 2D intensity images. The system combines contour-based and shading-based methods and uses a 3D tube model. These cylinders are defined by a cross-sectional element that is swept along the axis of the tube using some sweep rules. There are two main stages in the algorithm: local recognition stage and global recognition stage. The first step in the local recognition stage is the detection of reliable contour primitives. These primitives provide constraints for the localization of the tubes. Next, optimal filters are generated dynamically and matched against the data in order to verify the shading property of the tubes under detection. In the global recognition stage, locally verified tubes from the first stage are used as seeds and are swept along the axis of the tube using some sweep rule using best fit constraint. The key issue is to control the smoothness of the sweeping.
8 Conclusion
Segmentation algorithms form the essence of medical image applications such as radiological diagnostic systems, multimodal image registration, creating anatomical atlases, visualization, and computer-aided surgery. Even though many promising techniques and algorithms have been developed, it is still an open area for more research. The future direction of segmentation research will be towards developing faster and more accurate more automated techniques.
Fast advances in radiological imaging systems result in high volume patient images. Processing of these images in radiological diagnostic systems requires fast segmentation algorithms. One way to achieve faster segmentation results is to develop parallel algorithms. Cronemeyer et al [52] exploit the parallel nature of the hardware and develop a fast skeleton finder algorithm. Neural network-based approaches are also achieve faster segmentation due to their parallel nature. Another approach to achieve faster segmentation is to employ multiscale processing technique. In multi-scale image processing technique, major structures are extracted using low resolution images while fine structures are extracted using high resolution images. Accuracy of the segmentation process is crutial due to the nature of the work and is essential to achieve more precise and repeatable radiological diagnostic systems. Accuracy can be improved by incorporating apriori information on vessel anatomy and let high level knowledge guide the segmentation algorithm. Expert knowledge and guidence is essential in segmentation systems. While it is not expected that the segmentation systems will replace the experts, shear volume of the medical image data requires more automatic segmentation systems to reduce the work load.
We provide a survey of current vessel segmentation methods. We have tried to cover both early and recent literature related to vessel segmentation algorithms and techniques. Our aim was to introduce the current segmentation techniques. We intended to give the practitioner a framework for the existing research and to introduce interested parties to the panoply of vessel segmentation literature.
References
Owing to page constraints, the 52 references of this paper areavailableat http://vislab.cs.wright.edu/review/BIBE03-Ref.html. A larger version of the the Table 1 is also available at http://vislab.cs.wright.edu/review/BIBE03-Table.html.