Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Multimedia Systems (2019) 25:109–118

https://doi.org/10.1007/s00530-017-0580-7

SPECIAL ISSUE PAPER

Segmentation of blood vessels using rule-based and machine-


learning-based methods: a review
Fengjun Zhao1 · Yanrong Chen1 · Yuqing Hou1 · Xiaowei He1

Published online: 13 December 2017


© Springer-Verlag GmbH Germany, part of Springer Nature 2017

Abstract
Vessel segmentation as a component of medical image processing is the prerequisite for accurate diagnosis of vascular-related
diseases. Manual delineation of blood vessels has been turned out to be time consuming and observer dependent. Therefore,
much effort has been dedicated to the automatic or semi-automatic vessel segmentation methods. Previous literatures have
reviewed the state of vessel segmentation methods from various perspectives. However, their reviews did not take the mod-
ern machine-learning methods especially deep neural networks into account. In this paper, we reviewed the state-of-the-art
vessel segmentation methods by dividing them into two categories, rule-based, and machine-learning-based methods. The
rule-based methods discriminate vessel structure from background relying on intuitively and exquisitely designed rule sets,
while the machine-learning-based methods carry out the segmentation by self-learned rules from the previous experience.
Instead of exhaustively listing all vessel segmentation methods, this paper focuses on the well-known blood vessel segmen-
tation methods in recent years, to give readers a glimpse of the current state and future direction of segmentation technique
for blood vessels.

Keywords  Blood vessel · Segmentation · Rule-based · Machine learning · Deep neural network

1 Introduction growing number of automatic and semi-automatic (or inter-


active) vessel segmentation methods have been developed
The primary role of blood vessels in vertebrates is to sup- for the diagnosis of vascular-related diseases.
ply oxygen and nutrients to the surrounding cells and takes Several groups have reviewed the state of vessel segmen-
metabolic wastes away [1]. Vascular malfunctions can result tation methods from various perspectives. To the best of our
in many vascular-related diseases such as cardiac–cerebral knowledge, Kirbas and Quek [2] reviewed vessel extraction
vascular diseases, peripheral vascular diseases, and retinal techniques for the first time based on different mathematic
vascular disorders. Therefore, early diagnosis and precaution categories. Then, Lesage et al. [3] surveyed three-dimen-
of vascular-related diseases can greatly reduce the mortal- sional (3D) vessel lumen segmentation techniques from the
ity and disability. Vessel segmentation, as a component of viewpoints of models, features, and extraction schemes.
medical image processing, is the prerequisite of accurate These two literatures are important references for new prac-
diagnosis of vascular-related diseases. Manual delineation titioners just beginning with vessel segmentation. After-
by experts has been considered as golden standard, but it wards, Bibiloni et al. [4] surveyed the segmentation of curvi-
is time consuming and observer dependent. Therefore, a linear objects, including blood vessels as well as other thin,
long, line-like structures. Loizou [5] summarized the carotid
artery image and video segmentation in ultrasound imag-
Electronic supplementary material  The online version of this ing, which focused on the determination of intima–media
article (https://doi.org/10.1007/s00530-017-0580-7) contains thickness (IMT). The above literatures mainly reviewed
supplementary material, which is available to authorized users.
the methods based on presumed rules, which are designed
* Xiaowei He intuitively and exquisitely by experts. In other words, these
hexw@nwu.edu.cn segmentation methods are devised using the experience of
human beings rather than machines.
1
School of Information Sciences and Technology, Northwest
University, Xi’an 710069, Shaanxi, China

13
Vol.:(0123456789)

110 F. Zhao et al.

In addition, Fraz et al. [6] fully analyzed and catego- each pixel represents its affinity with vascular region. The
rized the blood vessel segmentation methodologies on reti- eigenvalue of a Hessian matrix can be exploited to form a
nal images. They mainly focused on the methods based on discriminant function, which has maximum responses for
machine learning. The procedure of vessel segmentation is structures like blood vessels. Similarly, Krissian et al. [11]
carried out by self-learned rules from the previous samples. utilized Hessian matrix method to obtain vessel centerlines,
These segmentation methods can be approximately divided and then reconstructed vessel walls with the information of
into supervised and unsupervised methods. In the first cat- vascular diameters. Zhou et al. [12] enhanced the coronary
egory, they systematically investigated vessel segmenta- artery by 3D multi-scale filtering and Hessian matrices, and
tion with artificial neural networks, support vector machine then tracked each coronary artery and identified the branches
(SVM), K-nearest neighbor, and Gaussian mixture model. In using 3D rolling balloon region growing.
the second category, they introduced the application of fuzzy The Hessian matrix method has two drawbacks. First,
C-means, Bayesian approaches, etc. However, the state-of- a Gaussian function is usually used to eliminate the influ-
the-art machine-learning methods have not yet been involved ence of noise on vesselness calculation. It applies the same
in vessel segmentation by then. Recently, large amount of scale on both blood vessels and background, leading to dis-
novel supervised or unsupervised vessel segmentation meth- turbance from adjacent structure. Hence, Xiao et al. [13]
ods have been proposed, among which the deep learning- replaced the Gaussian function by bi-Gaussian filtering,
based methods gain enormous success in the latest 5 years which allows independent selection of scale in blood ves-
[7]. Hence, it is of great importance to incorporate these up- sels and in background. Second, the Hessian matrix method
to-date methods into the community of vessel segmentation. is sensitive to gradient. Therefore, the sharp boundaries
In this paper, we classify the state-of-the-art vessel seg- between objective and background will affect the seg-
mentation methods into two major categories, rule-based mentation of blood vessels. To handle this problem, Zhao
methods, and machine-learning-based methods. The rule- et al. [14] patched these sharp boundaries from the region
based methods discriminate vessel structure from back- of interest using a smooth operation according to the sur-
ground relying on the intuitively and exquisitely designed rounding background voxels. This strategy finely remedies
rule sets, while the machine-learning-based methods carry the false-positive response of Hessian matrix methods on
out the segmentation by self-learned rules from the previous sharp boundaries, but at the expense of increased computa-
experience. Then, we further subdivide each category into tion time.
subclasses mainly based on the literatures from Web of Sci-
ence in recent years (Fig. S1 in Supplementary materials). 2.2 Marching filtering method
Instead of exhaustively listing all the vessel segmentation
methods, this paper focuses on the well-known methods in Marching filtering method assumes that the intensity along
recent years. the cross section of a blood vessel has Gaussian-like profile.
Ricci and Perfetti invented the line detector as a marching
filter for the segmentation of retinal blood vessels [15]. Line
2 Rule‑based method detector evaluates the average intensity along lines of fixed
length through the target pixel from different orientations.
In this category, numerous classification strategies have been Moreover, Nguyen et al. [16] combined the line detector
proposed. Their principle of classification varies, such as with multi-scale techniques by altering the length of a basic
mathematic categories, the procedure of vessel extraction, line detector. The final segmentation is attained through
and the degree of manual interaction. Nowadays, the most the linear combination of line responses at various scales.
commonly used classifications are based on rules consisting The marching filters with line detectors take original blood
of all the above principles. In a similar way of [8], we subdi- vessel images as inputs, but sometimes fail on images with
vide the most cited rule-based methods into Hessian matrix, high micro-vessel density and background noise. Thus,
marching filtering, mathematical morphology, minimal path, Ding et al. [17] presented an integrated filter for 3D vessel
active contour, and graph-based methods (Fig. S1 in Sup- segmentation. Their method incorporates the advantages of
plementary materials). multi-scale line detectors and Hessian matrix filters, which
performs well in the segmentation of micro-vessels.
2.1 Hessian matrix method Hessian matrix and line detector approaches have a com-
mon limitation. They make an implicit assumption that only
Blood vessels are generally thin, elongating, and bright single tubular structures are locally presented, which pre-
compared with the surrounding background. Sato et al. and cludes the application on complex blood vessels [9, 10, 16].
Frangi et al. [9, 10] enhanced the image of blood vessels Qian et al. [18] proposed a segmentation method to relax
using multi-scale vesselness method, in which the value of the assumption of single tubular structures. They captured

13
Segmentation of blood vessels using rule-based and machine-learning-based methods: a review 111

common properties shared by polar neighborhood intensity pixel classification was performed according to the fusion of
profiles in all types of vascular points. The method improves these resulting features.
the performance on vascular bifurcations based on the neigh- An apparent advantage of complex mathematical mor-
borhood intensity profile in spherical polar coordinates. phological methods over the basic ones lies in their capacity
However, it is sensitive to structural noise in proximity to of self-adaptive direction and scale recognition, benefited
background structures. Therefore, Rivest-Henault et al. [19] from the exquisite design of SEs. Nevertheless, the above
defined a 3D curvilinear structure detection filter based on complex morphological methods generally fail on microvas-
structure-ball (SBall) analysis. The SBall is constructed by culature when conducting direction estimation. Therefore,
evaluating second-order numerical differences in various there should be further investigation on accurate direction
directions. Since it is not specific to any specific distribution estimation of multi-scale blood vessels.
or profile of blood vessels, it can depict crossing structures
and junctions.
Except for using single feature in spatial domain, the 2.4 Minimal path method
marching filter can also be designed in multiple ways. Zhang
et al. [20] employed a pair of filters, the zero-mean Gaussian Minimal path techniques are commonly employed in inter-
filter, and the first-order derivative of Gaussian, to detect active frameworks. The method captures the global optimal
blood vessels. Lajevardi et al. [21] extracted blood vessels curve of a contour-dependent energy formulation between
using a family of matched filters in frequency domain, where user-supplied start and endpoints [32]. Minimal path meth-
the templates for blood vessels are defined as formal spatial ods have the superiority of global minimization, high com-
graphs derived from the retinal vascular network. Moreover, putational efficiency, and simple incorporation of user infor-
Kovacs et al. [22] proposed a novel marching template based mation. Wang et al. [33] proposed a vessel segmentation
on generalized Gabor function. They reconstructed the con- method that requires only a few mouse clicks. The method
tours of blood vessels using the extracted centerlines and the seeks an optimal path between given seed points by mini-
prior knowledge acquired from training database. mizing a cost function. Li and Yezzi [34] invented a four
dimensional (4D) minimal path method for vessel extraction.
2.3 Mathematical morphology method They modeled each blood vessel as a 4D curves, consisting
of three spatial coordinates and an extra vascular diameter
Traditional mathematical morphology methods utilize struc- dimension. In addition, Zhu et al. [44] segmented 3D coro-
ture elements (SEs) for medical image analysis, which rely nary arteries by minimizing the average edge cost along the
on basic operations, such as erosion, dilation, opening, and path in discrete 4D graph. It can simultaneously extract the
closing [23]. These mathematical morphology operations centerline and radius on the cross section of the coronary
are often used as a pre-processing in the extraction of tubu- artery without any image pre-processing. The limitation of
lar structures [24]. Subsequently, more complex morphol- these minimal path methods is that they do not consider ves-
ogy methods have been developed, including watersheds, sel orientation during vessel tracking.
gray-level hit-or-miss transform, and connective filter, which Therefore, Mohan et al. [35] associated the 4D path to
have been involved in 3D vessel filtering and segmentation anisotropic potentials, which exploited the orientation of
[25–27]. blood vessels. Afterwards, Benmansaour and Cohen [36]
Recently, Babin et al. extended the morphological pro- incorporated the information of vascular orientations by an
file by introducing more complex ring-shaped SEs [28], and anisotropic metric based on the optimally oriented flux. The
further defined the novel orientation-based line-shaped SEs metric is well-oriented along the direction of blood vessels,
[29]. Hereon, the line-shaped SEs incorporate the idea of allowing higher velocity on the vascular centerline and pro-
multi-scale approach into a multi-orientation morphology viding a good estimate of the vessel radius. Cetin et al. [37]
method, improving the computational efficiency of vessel viewed the vessel segmentation problem from the perspec-
segmentation. Moreover, Dufour et al. [30] performed a spa- tive of tensor estimation and tractography as in diffusion
tial variant morphological closing operation on 3D cerebral tensor imaging. The model can capture the whole vessel
vascular images based on the corrected directional field. trees through automatic branch detection algorithms with
The initial directional field calculated via Hessian matrix is single seed point.
refined using mathematical thinning and dilating operation, The minimal path method performs well on the blood
guided by the thresholded metrics of Hessian matrix. Most vessel confined by the user-supplied seed points, which
recently, Sigurosson et al. [31] combined directional math- has been widely used in vessel segmentation especially
ematical morphology with fuzzy classification. After feature for large vessels, such as aorta and carotid artery. How-
extraction through advanced morphological directional filter ever, the extraction of complex vascular networks like
based on the contrast of vessels and their linear connectivity, coronary arterial networks and cerebral blood vessels is

13

112 F. Zhao et al.

still troublesome, since the user-supplied seeds can hardly 2.6 Graph‑based method
cover the entire networks.
Graph-based methods, especially graph cuts, transfer the prob-
lem of global optimal segmentation into energy minimization
2.5 Active contour method [48]. These methods have been applied to multi-organ segmen-
tation from CT imaging [49]. However, the graph cut method
Based on different energy formulation (implicit or generally fails to extract blood vessels, since the energy mini-
explicit), active contour methods can be broadly divided mization tends to shrink for curvilinear structures. To handle
into two subclasses: level set method and snake-based this problem, the graph-based method needs prior knowledge
method. Developed by Chan and Vese [38], level set or the collaboration with other pre-processing techniques [50].
method converts the minimal partition problem into the Bauer et al. [51] separated multiple blood vessel trees
implicit energy minimization. The method has gained effectively using shape priors and graph cut method. First,
much attention in both natural and medical image seg- tree structures are formed in a top–down fashion, which uses
mentation [39, 40]. For example, Lorigo et  al. [41] knowledge about the blood direction in tubular networks
modeled the vessel boundary as a manifold that evolves based on centerline extraction. Second, these tree structures
iteratively to minimize an energy criterion. Manniesing are incorporated as a shape prior for constrained graph cut
et al. [42] improved the level set segmentation of vascular segmentation. Esneault et al. [52] combined the graph cut
structures with image enhancement methods by combin- method with a geometrical moment-based detector of cylin-
ing a smooth vessel filter based on a geometrical analysis drical shapes within the min-cut/max-flow energy minimiza-
of the Hessian eigenvalue. However, these level set-based tion framework. This method automates the segmentation
methods lead to unwanted leakage to background regions by adding a data term into the widely used Boykov’s graph
because of the absence of constraints during function evo- cuts algorithm.
lution. Therefore, Manniesing et al. [43] later presented On the other hand, Zhao et al. [53] segmented the retinal
an 3D vessel tracking approach, which guided the evolu- vasculature in three steps: image inhomogeneity correction,
tion of the surface by analyzing its skeleton topology and vessel enhancement, and graph-based contour segmentation.
imposing shape constraints. Forkert et al. [44] integrated The first two steps as pre-processing enhance the image of
an additional vesselness force in the level set formulation. retinal blood vessels and facilitate the subsequent segmen-
This method weights the internal energy lower if the gra- tation. Hernandez-Vela et al. [54] combined vesselness,
dient direction of the level set is similar to the direction geodesic paths, and a multi-scale edgeness map with graph
of the vesselness eigenvector. cuts approach, which proceeded vessel segmentation by
Depending on explicit Lagrangian contour evolutional global optimization of the graph cuts energy function. Roy-
formulations, snake has become one of the most active chowdhury et al. [55] presented a graph-based method that
research areas in medical image segmentation [45]. Xu exploited a new adaptive cylinder flux (ACF)-based connec-
et al. [46] applied snake model to detect the boundary tivity framework, which was formulated based on random
of intima–media for IMT measurement, in which two walks. With the help of ACF-based connectivity framework,
contours are initialized from line segments generated by the global optimal segmentation can be achieved without any
Hough transform, and then, dual snake model is evolved prior knowledge.
simultaneously to detect boundary. Moreover, Yuanzhi Couple with prior knowledge and pre-processing, graph-
et al. [47] elaborately designed a segmentation method for based method has been successfully adopted to the area of
thin blood vessels with low contrast through the introduc- vessel segmentation. This method models the segmentation
tion of a constrained B-snake. This method starts with the of vessel from background as finding the best cut across a
extraction of vessel axis from the original CT angiogram, graph, which is solved by optimization. However, the graph-
followed by the shape- and size-constrained snake-based based method is vulnerable to pre-processing, image quality,
vessel boundary delineation on cross sections. and even optimization parameters. Thus, suitable denois-
The active contour method can be implemented either ing, pre-processing, and even parameter tuning are required
automatically or interactively for vessel segmentation. In before the utilization of graph-based method.
general, this method performs worse on the segmentation
of blood vessels than on the segmentation of regional
organs, such as livers and lungs. This is because this 3 Machine‑learning‑based method
method is sensitive to the initialization, which results in
the poor accuracy for the thin and elongated structure of Machine-learning-based method puts vessel segmentation as
blood vessels. the problem of pixel classification by assigning each pixel as
vessel or non-vessel. Pixel classification generally consists

13
Segmentation of blood vessels using rule-based and machine-learning-based methods: a review 113

of training and testing. In the training stage, the method Fuzzy C-means clustering to classify retinal vessels based
learns the correct classification of pixels from known labels on the textural feature extraction. Kande et al. [64] enhanced
(ground truth), and then, in test stage, the trained classifier is the contrast of blood vessels using matching filtering, and
examined on previously unseen pixels. This machine-learn- then segmented vascular tree from retinal images by spa-
ing-based method is the typical supervised method, while tially weighted Fuzzy C-means clustering. Clustering meth-
some other machine-learning methods are unsupervised, ods, such as K-means, can also be used as pre-processing
such as K-means clustering, which learns the features for to reduce the negative impact of bright lesions in fundus
classification themselves without given labels in the training images, followed by vessel extraction with rule-based meth-
stage. Thus, both the supervised and unsupervised vessel ods [65].
segmentation methods are categorized as machine-learning- Support vector machines (SVM) is one of the most distin-
based methods. We further divide machine-learning-based guished machine-learning-based methods for medical image
methods into the conventional machine-learning, decision- segmentation for its well-justified theoretical basis [66].
tree-based method, and deep neural networks (Fig. S1 in When applied to vessel extraction, SVM has to be collabo-
Supplementary materials). rated with other feature extraction approaches, e.g., march-
ing filtering [15]. Xu et al. [67] utilized a two-step method
3.1 Conventional machine learning for retinal vessel segmentation to overcome the variations
in the contrast of large and thin vessels. They first extracted
The basic function of artificial neural networks (NNs) is to large vessel by adaptive local thresholding, and then adopted
learn, remember, and then generalize the data to produce SVM classification and tracking growth to segment the thin
new information [56]. Gardner et al. [57] applied a back blood vessels. Afterwards, You et al. [68] designed a scheme
propagation multilayer NNs to retinal vascular tree segmen- based on radial projection and SVM method to extract both
tation. For efficient computation, Sinthanayothin et al. [58] the large and the thin retinal blood vessels. Radial projection
used a multilayer perceptron NNs for vessel detection. Each is used to locate vessel centerlines from narrow vessels with
pixel is classified using the first principal component after low contrast, followed by vessel enhancement through modi-
principal components analysis (PCA), combining with the fication of a steerable complex wavelet. Then, the vessel
edge strength values from the patch centered on the current pixel is presented by a feature vector through line strength,
pixel. Recently, Diego et al. [8] suggested the feed-forward where the vector is provided for SVM training. Besides,
multilayer NNs based on a 7D vector, consisting of gray- Agurto et al. [69] presented a multi-scale method to detect
level and moment-invariant features. Although NNs have neovascularization in the optic discs. Actually, because of
gained much attention in various fields, they are seldom used its ability of local pattern recognition, textural features have
in the segmentation of blood vessels. In fact, accurate weight great potential complements to intensity-based features for
calculation for NNs requires large amount of training data, vessel segmentation [70].
which exceeds the computational capability of most personal
computers before the popularization of parallel computing. 3.2 Decision‑tree‑based method
K-nearest neighbor can also be used to delineate blood
vessels in advantage of its simplicity [59]. For example, Built from the whole data set covering all features, decision-
Staal et al. [60] presented a ridge-based vessel segmenta- tree algorithms are effective for providing human-readable
tion method with a K-nearest neighbor classifier in retinal rules of classification [71]. However, they are vulnerable
images. Moreover, Gaussian mixture model is another con- to small perturbations in the learning data set [72]. It has
ventional machine-learning-based method used for vessel been demonstrated that bagging techniques can effectively
segmentation. It assumes that the histograms of blood ves- alleviate this problem [73]. As an ensemble of decision trees,
sels and its surrounding backgrounds comply with multiple random forest combines random subspace method and bag-
Gaussian distributions [61]. Recently, Roychowdhury et al. ging techniques, and has become a popular technique for
[62] suggested a three-stage blood vessel segmentation algo- classification, prediction, and variable selection [74]. Both
rithm using fundus photograph. They first pre-processed the decision trees and random forests have been brought into the
green plane of a fundus image to extract major vessels, then segmentation of blood vessel.
classified all the remaining pixels using a Gaussian mix- Fraz et  al. [75] segmented retinal blood vessel from
ture model classifier with a set of eight features, and finally retinal photograph using an ensemble system of bagged
combined the major portions of the blood vessels with the and boosted decision trees. Subsequently, they applied the
classified vessel pixels. boosted decision-tree method into the retinal vessel segmen-
The K-means and Fuzzy C-means clustering as typical tation in multi-ethnic school children image with a different
unsupervised machine-learning methods, are used in reti- feature vector [76]. It is demonstrated that the ensemble-
nal vessel extraction as well. Bhuiyan et al. [63] applied based method provides a tool for clinical large population

13

114 F. Zhao et al.

studies. Furthermore, Shahrian et al. [77] constructed a clas- To the best of our knowledge, Liskowski et  al. [86]
sifier based on ensembles of n decision trees to segment published the first paper that genuinely applied the deep
vessels in color retinal images. They reduced the data dimen- convolutional neural network to the segmentation of reti-
sion and increased the discriminability of the feature sets by nal blood vessels. They exhaustively introduced the CNNs
PCA and linear discriminant analysis (LDA), respectively. form the view of local connectivity, parameter sharing
The bagging technique is implemented to provide independ- and pooling, as well as the combination with the problem
ent training and test image sets, which overcomes the prob- of vessel segmentation. Since then, the CNNs have been
lem of overfitting. fully exploited in numerous vessel extraction applications.
Aslani et al. [78] adopted a random forest vessel seg- Khalaf et al. [87] modified the segmentation of retinal
mentation method using 17 features for its superiority in vessels to reduce the intra-class variance by formulating
speed, simplicity, and information fusion. The combination the problem as a three-class problem that differentiates
of individually strong features leads to a better discrimi- large vessels, small vessels, and background areas. They
nation between vessel and non-vessel pixels. Cherry et al. found that the combination of convolutional kernels with
[79] suggested a marginal artery localization method from different sizes in CNNs achieved the best sensitivity and
computed tomography colonography using sequential Monte specificity. Wu et al. [88] trained a deep CNN to estimate
Carlo tracking by random forest regression. The random for- the local vessel probability via PCA and nearest neigh-
est algorithm was employed for intelligent cue fusion and bor search, and fed the resulting map into a probabilistic
decision making, which achieved high sensitivity and robust- tracking framework to extract the entire connected ves-
ness. Moreover, Schneider et al. [80] developed a framework sel trees. Similarly, Prentasic et al. [89] applied CNNs
for 3D vessel segmentation and centerline extraction based to vessel segmentation in optical coherence tomography
on oblique random forests and Hough voting. It relies on a angiography (OCT-A) images. With manual segmentation
steerable filter for the efficient computation of local image as the ground truth, the results demonstrate high effective-
features at different scale and orientation. Until now, we find ness of CNNs in OCT-A images. Moreover, Maji et al. [90]
that almost all the random forests based vessel segmentation suggested an ensemble of CNNs for vessel segmentation
methods have to deal with handcrafted feature extraction, in fundus color images. They first trained an ensemble of
which is labor intensive and time consuming. Therefore, deep convolutional neural networks to segment vessel and
Wang et al. [81] tackled this problem using a superior clas- non-vessel areas, and then averaged the responses of the
sifier, i.e., the convolutional NN. Convolutional NN (that individual CNNs.
will be descripted in Sect. 3.3), performing as a trainable Most recently, Li et al. [91] viewed the vessel segmen-
hierarchical feature extractor, can automatically learn fea- tation as an issue of cross-modality data transformation
tures from the raw images and greatly save the time spent from retinal image to vessel map. A wide and deep neu-
on handcrafted feature extraction. ral network with strong induction ability is required to
model the relation between the retinal image and vessel
3.3 Deep neural networks map. This method reduces the impact of subjective fac-
tors, since it needs no pre-processing, but at the cost of
With the development of computation capability, deep neu- increased computation. Furthermore, Moeskops et al. [92]
ral networks (NNs) have won numerous contests in image investigated the transfer learning capacity of a single CNN
processing and machine learning, Deep NNs allow com- on different segmentation tasks. They demonstrated that
putational models composed of multiple processing layers the segmentation performance of a CNN trained specifi-
to learn representative knowledge of data through multiple cally for each task was equivalent to that of the combined
levels of abstraction [82]. Compared with conventional NNs, training of three tasks (brain MRI, breast MRI, and car-
deep neural networks can effectively and robustly train net- diac CTA). This means that a trained CNN on the tissue
works using efficient activation functions and dropout strat- segmentation can be potentially used in coronary artery
egy [83, 84]. Amongst a variety of deep NNs, the most suc- segmentation without vessel-specific training. In addition,
cessful type of models for image analysis is convolutional Annunziata et al. [93] observed that the filters learned by
neural networks (CNNs). The CNNs have two advantages: the lower layers of convolutional sparse coding (CSC)
first, image features are automatically learned instead of for curvilinear structure segmentation were similar to the
handcrafted feature extraction as in conventional machine handcrafted filters. Thus, they proposed an efficient ves-
learning; second, the procedure of automatic feature extrac- sel segmentation approach by combining the CSC filters
tion with multiple layers is hierarchical from simple to com- with handcrafted filters, which achieved high speed and
plex. Therefore, the CNNs have drawn much attention in discriminative power simultaneously.
numerous applications, especially in medical image analysis
[85].

13
Segmentation of blood vessels using rule-based and machine-learning-based methods: a review 115

4 Conclusions and discussion DRIVE and STARE databases. However, these deep learn-
ing methods are seldom involved in the segmentation of cer-
In this paper, we have reviewed the most commonly used ebral blood vessel and coronary artery, likely due to the lack
vessel segmentation methods in recent years. The machine- of standard labeled databases. The requirement of massive
learning-based methods are usually more computational data is a characteristic of machine-learning-based methods,
expensive than the rule-based methods, because the training and the more data we collect, the more robust classifier can
involves large amount of data. However, the machine-learn- we train. Hence, the adoption of deep neural networks into
ing-based methods, especially deep neural networks, take other vessel segmentation tasks heavily depends on the con-
raw medical images as input without manual feature extrac- struction of the standard database. In addition, users may
tion, thus dramatically improve the accuracy and robustness also have to participate in the labeling of database with inter-
of vessel segmentation. The detailed assessment of different active techniques, in which the label is subsequently checked
vessel segmentation methods can be found in Supplementary and modified by experts.
materials. As a summary, the conclusions and trends of the
latest vessel segmentation methods are described as the fol- Acknowledgements  This work was supported in part by the National
Natural Science Foundation of China under Grant Nos. 61601363,
lowing aspects. 61372046, 61640418, 61401264, 11571012, 81530058, 61601154,
and 61502387, the National Key R&D Program of China under Grant
No. 2016YFC1300300, the Science and Technology Plan Program
4.1 Vessel segmentation is closely related in Shaanxi Province of China under Grant Nos. 2013K12-20-12 and
to centerline extraction 2015KW-002, the Natural Science Research Plan Program in Shaanxi
Province of China under Grant Nos. 2017JQ6017, 2015JM6322, and
2015JZ019, and the Scientific Research Foundation of Northwest
Vessel centerline extraction is equally important as vessel
University.
segmentation for the topological analysis of vascular net-
works [94, 95]. In most case, centerline extraction is con-
Compliance with ethical standards 
ducted on binary images after vessel segmentation [96–98].
However, it is synchronized with the vessel segmentation in Conflict of interest  All authors declare that they have no conflict of
some applications [99]. Among different methods, minimal interest in our research.
path vessel segmentation method is most likely to be accom-
panied with centerline extraction, which is manifested by the
path tracking between the given seed points in blood vessels.
References
4.2 Hybrid method gradually replaces the method 1. Lorthois, S., Cassot, F.: Fractal analysis of vascular networks:
with single technique insights from morphogenesis. J. Theor. Biol. 262(4), 614–633
(2010)
Hybrid methods have been increasingly adopted to build 2. Kirbas, C., Quek, F.: A review of vessel extraction techniques
and algorithms. ACM Comput. Surv. 36(2), 81–121 (2004)
discriminative vessel detection method. Typical hybrid 3. Lesage, D., Angelini, E.D., Bloch, I., et al.: A review of 3D vessel
methods are the combination within rule-based methods, lumen segmentation techniques: models, features and extraction
and between rule-based and machine-learning-based meth- schemes. Med. Image Anal. 13(6), 819–845 (2009)
ods. The first type of combination is a powerful tool used 4. Bibiloni, P., Gonzalez-Hidalgo, M., Massanet, S.: A survey on
curvilinear object segmentation in multiple applications. Pattern
for feature extraction, but requires exquisite design and is Recogn. 60, 949–970 (2016)
only suitable for a specific type of medical images [42, 52, 5. Loizou, C.P.: A review of ultrasound common carotid artery
54]. The second type combines the advantages of fast mul- image and video segmentation techniques. Med. Biol. Eng. Com-
tiple feature extraction in rule-based methods and the robust put. 52(12), 1073–1093 (2014)
6. Fraz, M.M., Remagnino, P., Hoppe, A., et al.: Blood vessel seg-
classifier in machine-learning-based methods [31, 78, 100]. mentation methodologies in retinal images—a survey. Comput.
Methods Progr. Biomed. 108(1), 407–433 (2012)
7. Litjens, G., Kooi, T., Bejnordi, B.E., et al.: A survey on deep
4.3 Deep learning will be fully exploited in vessel learning in medical image analysis. Med. Image Anal.  42,
segmentation 60–88 (2017)
8. Marin, D., Aquino, A., Emilio Gegundez-Arias, M., et al.: A
Vessel segmentation with deep neural networks has shown new supervised method for blood vessel segmentation in retinal
images by using gray-level and moment invariants-based fea-
great superiority over rule-based methods and conven- tures. IEEE Trans. Med. Imaging 30(1), 146–158 (2011)
tional neural networks. Nowadays, we find that the CNNs 9. Sato, Y., Nakajima, S., Atsumi, H., et al.: 3D multi-scale line
come into the application of retinal vessel segmentation filter for segmentation and visualization of curvilinear struc-
and achieve huge success in two public databases, i.e., the tures in medical images. In: Cvrmed-Mrcas’97: First joint
conference—Computer vision, virtual reality and robotics in

13

116 F. Zhao et al.

medicine and medical robotics and computer-assisted surgery, arteriovenous malformation segmentation. Med. Image Anal.
vol. 1205, pp 213–222 (1997) 16(5), 991–1002 (2012)
10. Frangi, A.F., Niessen, W.J., Vincken, K.L., et al.: (1998) Mul- 29. Babin, D., Pizurica, A., De Vylder, J., et al.: Brain blood ves-
tiscale vessel enhancement filtering. In: Lecture Notes in Com- sel segmentation using line-shaped profiles. Phys. Med. Biol.
puter Science. In: Proceedings of Medical Image Computing 58(22), 8041–8061 (2013)
and Computer-Assisted Intervention—MICCAI’98, 1998 1998. 30. Dufour, A., Tankyevych, O., Naegel, B., et al.: Filtering and seg-
Springer, Berlin, Germany, pp 130–137 mentation of 3D angiographic data: advances based on math-
11. Krissian, K., Malandain, G., Ayache, N., et al.: Model-based ematical morphology. Med. Image Anal. 17(2), 147–164 (2013)
detection of tubular structures in 3D images. Comput. Vis. 31. Sigurosson, E.M., Valero, S., Benediktsson, J.A., et al.: Auto-
Image Underst. 80(2), 130–171 (2000) matic retinal vessel extraction based on directional mathematical
12. Zhou, C., Chan, H.P., Chughtai, A., et al.: Computerized analy- morphology and fuzzy classification. Pattern Recogn. Lett. 47,
sis of coronary artery disease: performance evaluation of seg- 164–171 (2014)
mentation and tracking of coronary arteries in CT angiograms. 32. Cohen, L.D., Kimmel, R.: Global minimum for active contour
Med. Phys. 41(8), 081912 (2014) models: a minimal path approach. Int. J. Comput. Vis. 24(1),
13. Xiao, C., Staring, M., Wang, Y., et al.: Multiscale Bi-Gaussian 57–78 (1997)
filter for adjacent curvilinear structures detection with applica- 33. Wang, L., Kallem, V., Bansal, M., et al.: Interactive retinal ves-
tion to vasculature images. IEEE Trans. Image Process. 22(1), sel extraction by integrating vessel tracing and graph search. In:
174–188 (2013) Medical image computing and computer-assisted intervention:
14. Zhao, F., Liang, J., Chen, D., et al.: Automatic segmentation MICCAI International Conference on Medical Image Comput-
method for bone and blood vessel in murine hindlimb. Med. ing and Computer-Assisted Intervention 16 (Pt 2), pp. 567–574
Phys. 42(7), 4043–4054 (2015) (2013)
15. Ricci, E., Perfetti, R.: Retinal blood vessel segmentation using 34. Li, H., Yezzi, A.: Vessels as 4-d curves: global minimal 4-d paths
line operators and support vector classification. IEEE Trans. to extract 3-d tubular surfaces and centerlines. IEEE Trans. Med.
Med. Imaging 26(10), 1357–1365 (2007) Imaging 26(9), 1213–1223 (2007)
16. Nguyen, U.T.V., Bhuiyan, A., Park, L.A.F., et al.: An effective 35. Mohan, V., Sundaramoorthi, G., Tannenbaum, A.: Tubular sur-
retinal blood vessel segmentation method using multi-scale face segmentation for extracting anatomical structures from
line detection. Pattern Recogn. 46(3), 703–715 (2013) medical imagery. IEEE Trans. Med. Imaging 29(12), 1945–1958
17. Ding, Y., Ward, W.O.C., Waesterlid, T., et al.: Three-dimen- (2010)
sional vessel segmentation using a novel combinatory filter 36. Benmansour, F., Cohen, L.D.: Tubular structure segmentation
framework. Phys. Med. Biol. 59(22), 7013–7029 (2014) based on minimal path method and anisotropic enhancement.
18. Qian, X., Brennan, M.P., Dione, D.P., et al.: A non-parametric Int. J. Comput. Vis. 92(2), 192–210 (2011)
vessel detection method for complex vascular structures. Med. 37. Cetin, S., Demir, A., Yezzi, A., et al.: Vessel tractography using
Image Anal. 13(1), 49–61 (2009) an intensity based tensor model with branch detection. IEEE
19. Rivest-Henault, D., Cheriet, M.: 3-D curvilinear structure Trans. Med. Imaging 32(2), 348–363 (2013)
detection filter via structure-ball analysis. IEEE Trans. Image 38. Chan, T.F., Vese, L.A.: Active contours without edges. IEEE
Process. 22(7), 2849–2863 (2013) Trans. Image Process. 10(2), 266–277 (2001)
20. Zhang, B., Zhang, L., Zhang, L., et al.: Retinal vessel extrac- 39. Li, C., Huang, R., Ding, Z., et al.: A level set method for image
tion by matched filter with first-order derivative of Gaussian. segmentation in the presence of intensity inhomogeneities with
Comput. Biol. Med. 40(4), 438–445 (2010) application to MRI. IEEE Trans. Image Process. 20(7), 2007–
21. Lajevardi, S.M., Arakala, A., Davis, S.A., et al.: Retina verifi- 2016 (2011)
cation system based on biometric graph matching. IEEE Trans. 40. Vese, L.A., Chan, T.F.: A multiphase level set framework for
Image Process. 22(9), 3625–3635 (2013) image segmentation using the Mumford and Shah model. Int. J.
22. Kovacs, G., Hajdu, A.: A self-calibrating approach for the seg- Comput. Vis. 50(3), 271–293 (2002)
mentation of retinal vessels by template matching and contour 41. Lorigo, L.M., Faugeras, O.D., Grimson, W.E.L., et al.: CURVES:
reconstruction. Med. Image Anal. 29, 24–46 (2016) curve evolution for vessel segmentation. Med. Image Anal. 5(3),
23. Najman, L., Talbot, H.: Introduction to mathematical mor- 195–206 (2001)
phology. Comput. Vis. Graph. Image Process. 35(3), 283–305 42. Manniesing, R., Viergever, M.A., Niessen, W.J.: Vessel enhanc-
(1986) ing diffusion—a scale space representation of vessel structures.
24. Zana, F., Klein, J.C.: Segmentation of vessel-like patterns using Med. Image Anal. 10(6), 815–825 (2006)
mathematical morphology and curvature evaluation. IEEE Trans. 43. Manniesing, R., Viergever, M.A., Niessen, W.J.: Vessel axis
Image Process. 10(7), 1010–1019 (2001) tracking using topology constrained surface evolution. IEEE
25. Passat, N., Ronse, C., Baruthio, J., et al.: Watershed and multi- Trans. Med. Imaging 26(3), 309–316 (2007)
modal data for brain vessel segmentation: application to the supe- 44. Forkert, N.D., Schmidt-Richberg, A., Fiehler, J., et al.: 3D cer-
rior sagittal sinus. Image Vis. Comput. 25(4), 512–521 (2007) ebrovascular segmentation combining fuzzy vessel enhancement
26. Bouraoui, B., Ronse, C., Baruthio, J., et al.: 3D segmentation of and level-sets with anisotropic energy weights. Magn. Reson.
coronary arteries based on advanced mathematical morphology Imaging 31(2), 262–271 (2013)
techniques. Comput. Med. Imaging Gr. 34(5), 377–387 (2010) 45. McInerney, T., Terzopoulos, D.: T-snakes: topology adaptive
27. Caldairou, B., Passat, N., Naegel, B.: (2010) Attribute-Filter- snakes. Med. Image Anal. 4(2), 73–91 (2000)
ing and Knowledge Extraction for Vessel Segmentation. In: 46. Xu, X., Zhou, Y., Cheng, X., et al.: Ultrasound intima-media
Bebis G, Boyle R, Parvin B et al. (eds) Advances in Visual segmentation using Hough transform and dual snake model.
Computing: 6th International Symposium, ISVC 2010, Las Comput. Med. Imaging Gr. 36(3), 248–258 (2012)
Vegas, NV, USA, November 29-December 1, 2010. Proceed- 47. Yuanzhi, C., Xin, H., Ji, W., et al.: Accurate vessel segmentation
ings, Part I. Springer, Berlin, Heidelberg, pp 13–22. https://doi. with constrained B-snake. IEEE Trans. Image Process. 24(8),
org/10.1007/978-3-642-17289-2_2 2440–2455 (2015)
28. Babin, D., Pizurica, A., Bellens, R., et al.: Generalized pixel 48. Boykov, Y., Funka-Lea, G.: Graph cuts and efficient N-D image
profiling and comparative segmentation with application to segmentation. Int. J. Comput. Vis. 70(2), 109–131 (2006)

13
Segmentation of blood vessels using rule-based and machine-learning-based methods: a review 117

49. Wolz, R., Chu, C., Misawa, K., et al.: Automated abdominal 70. Zhang, J., Liang, J., Zhao, H.: Local energy pattern for texture
multi-organ segmentation with subject-specific atlas generation. classification using self-adaptive quantization thresholds. IEEE
IEEE Trans. Med. Imaging 32(9), 1723–1730 (2013) Trans. Image Process. 22(1), 31–42 (2013)
50. Liu, M., Zhang, J., Yap, P.-T., et al.: View-aligned hypergraph 71. Ben-Haim, Y., Tom-Tov, E.: A streaming parallel decision tree
learning for Alzheimer’s disease diagnosis with incomplete algorithm. J. Mach. Learn. Res. 11, 849–872 (2010)
multi-modality data. Med. Image Anal. 36, 123–134 (2017) 72. Verikas, A., Gelzinis, A., Bacauskiene, M.: Mining data with
51. Bauer, C., Pock, T., Sorantin, E., et al.: Segmentation of inter- random forests: a survey and results of new tests. Pattern Rec-
woven 3d tubular tree structures utilizing shape priors and graph ogn. 44(2), 330–349 (2011)
cuts. Med. Image Anal. 14(2), 172–184 (2010) 73. Dietterich, T.G.: An experimental comparison of three methods
52. Esneault, S., Lafon, C., Dillenseger, J.-L.: Liver vessels segmen- for constructing ensembles of decision trees: bagging, boost-
tation using a hybrid geometrical moments/graph cuts method. ing, and randomization. Mach. Learn. 40(2), 139–157 (2000)
IEEE Trans. Biomed. Eng. 57(2), 276–283 (2010) 74. Biau, G., Devroye, L., Lugosi, G.: Consistency of random for-
53. Zhao, Y., Liu, Y., Wu, X., et al.: (2015) Retinal vessel segmenta- ests and other averaging classifiers. J. Mach. Learn. Res. 9,
tion: an efficient graph cut approach with retinex and local phase. 2015–2033 (2008)
Plos One 10 (4) 75. Fraz, M.M., Remagnino, P., Hoppe, A., et al.: An ensemble
54. Hernandez-Vela, A., Gatta, C., Escalera, S., et al.: Accurate classification-based approach applied to retinal blood vessel
coronary centerline extraction, caliber estimation, and catheter segmentation. IEEE Trans. Biomed. Eng. 59(9), 2538–2548
detection in angiographies. IEEE Trans. Inf. Technol. Biomed. (2012)
16(6), 1332–1340 (2012) 76. Fraz, M.M., Rudnicka, A.R., Owen, C.G., et al.: Delineation of
55. Grady, L.: Random walks for image segmentation. IEEE Trans. blood vessels in pediatric retinal images using decision trees-
Pattern Anal. Mach. Intell. 28(11), 1768–1783 (2006) based ensemble classification. Int. J. Comput. Assist. Radiol.
56. Yazici, A.C., Ogus, E., Ankarali, S., et al.: Artificial neural net- Surg. 9(5), 795–811 (2014)
works: review. Turk Klin Tip Bilim Derg 27(1), 65–71 (2007) 77. Shahrian, E.V., Yousefi, S., Isfahani, A.M., et al.: Vessels seg-
57. Gardner, G.G., Keating, D., Williamson, T.H., et al.: Automatic mentation in color retinal images using ensemble of bagged deci-
detection of diabetic retinopathy using an artificial neural net- sion trees and patched based principle component analysis and
work: a screening tool. Br. J. Ophthalmol. 80(11), 940–944 linear discriminant analysis. Investig. Ophthalmol. Vis. Sci. 56,
(1996) 7 (2015)
58. Sinthanayothin, C., Boyce, J.F., Cook, H.L., et al.: Automated 78. Aslani, S., Sarnel, H.: A new supervised retinal vessel segmen-
localisation of the optic disc, fovea, and retinal blood vessels tation method based on robust hybrid features. Biomed. Signal
from digital colour fundus images. Br. J. Ophthalmol. 83(8), Process. Control 30, 1–12 (2016)
902–910 (1999) 79. Cherry, K.M., Peplinski, B., Kim, L., et al.: Sequential Monte
59. Niemeijer, M., Staal, J., van Ginneken, B., et al.: Comparative Carlo tracking of the marginal artery by multiple cue fusion and
study of retinal vessel segmentation methods on a new publicly random forest regression. Med. Image Anal. 19(1), 164–175
available database. In: Proceedings of SPIE, pp. 648–656 (2004) (2015)
60. Staal, J., Abramoff, M.D., Niemeijer, M., et al.: Ridge-based ves- 80. Schneider, M., Hirsch, S., Weber, B., et al.: Joint 3-D vessel seg-
sel segmentation in color images of the retina. IEEE Trans. Med. mentation and centerline extraction using oblique Hough forests
Imaging 23(4), 501–509 (2004) with steerable filters. Med. Image Anal. 19(1), 220–249 (2015)
61. Soares, J.V.B., Leandro, J.J.G., Cesar, R.M. Jr., et al.: Retinal 81. Wang, S., Yin, Y., Cao, G., et al.: Hierarchical retinal blood ves-
vessel segmentation using the 2-D Gabor wavelet and supervised sel segmentation based on feature and ensemble learning. Neu-
classification. IEEE Trans. Med. Imaging 25(9), 1214–1222 rocomputing 149, 708–717 (2015)
(2006) 82. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature
62. Roychowdhury, S., Koozekanani, D.D., Parhi, K.K.: Blood Ves- 521(7553), 436–444 (2015)
sel Segmentation of Fundus Images by Major Vessel Extraction 83. Glorot, X., Bordes, A., Bengio, Y.: Deep Sparse Rectifier Neural
and Subimage Classification. IEEE J. Biomed. Health Inf. 19(3), Networks. In: International Conference on Artificial Intelligence
1118–1128 (2015) and Statistics (AISTATS), 11–13 April 2011, pp. 315–323 (2011)
63. Bhuiyan, A., Nath, B., Chua, J., et al.: Blood vessel segmentation 84. Srivastava, N., Hinton, G., Krizhevsky, A., et al.: Dropout: a
from color retinal images using unsupervised texture classifica- simple way to prevent neural networks from overfitting. J. Mach.
tion. In: 2007 IEEE International Conference on Image Process- Learn. Res. 15, 1929–1958 (2014)
ing, Sept. 16 2007–Oct. 19 2007, pp. V-521–V-524 (2007) 85. Zhang, J., Liu, M., Shen, D.: Detecting anatomical landmarks
64. Kande, G.B., Subbaiah, P.V., Savithri, T.S.: Unsupervised fuzzy from limited medical imaging data using two-stage task-oriented
based vessel segmentation in pathological digital fundus images. deep neural networks. IEEE Trans. Image Process. 26(10), 4753–
J. Med. Syst. 34(5), 849–858 (2010) 4764 (2017)
65. Saffarzadeh, V.M., Osareh, A., Shadgar, B.: Vessel Segmentation 86. Liskowski, P., Krawiec, K.: Segmenting retinal blood vessels
in retinal images using multi-scale line operator and K-Means with deep neural networks. IEEE Trans. Med. Imaging 35(11),
clustering. J.Med. Sign. Sens. 4(2), 122–129 (2014) 2369–2380 (2016)
66. Zhang, J., Liu, M., An, L., et al.: Alzheimer’s disease diagnosis 87. Khalaf, A.F., Yassine, I.A., Fahmy A.S.: Convolutional neural
using landmark-based features from longitudinal structural MR networks for deep feature learning in retinal vessel segmentation.
images. IEEE J. Biomed. Health Inf. 21(6), 1607–1616 (2017) In: 2016 IEEE International Conference on Image Processing
67. Xu, L., Luo, S.: A novel method for blood vessel detection from (ICIP), 25–28 Sept. 2016, pp. 385–388 (2016)
retinal images. Biomed. Eng. Online 9, 14 (2010) 88. Wu, A., Xu, Z., Gao, M., et al.: Deep vessel tracking: A general-
68. You, X., Peng, Q., Yuan, Y., et al.: Segmentation of retinal blood ized probabilistic approach via deep learning. In: 2016 IEEE 13th
vessels using the radial projection and semi-supervised approach. International Symposium on Biomedical Imaging (ISBI), 13–16
Pattern Recogn. 44(10–11), 2314–2324 (2011) April 2016, pp. 1363–1367 (2016)
69. Agurto, C., Yu, H., Murray, V., et al.: A multiscale decomposi- 89. Prentasic, P., Heisler, M., Mammo, Z., et al.: (2016) Segmenta-
tion approach to detect abnormal vasculature in the optic disc. tion of the foveal microvasculature using deep learning networks.
Comput. Med. Imaging Gr. 43, 137–149 (2015) J. Biomed. Opt. 21, 7

13

118 F. Zhao et al.

90. Maji, D., Santara, A., Mitra, P., et al.: (2016) Ensemble of deep 95. Zhao, F., Liang, J., Chen, X., et al.: Quantitative analysis of
convolutional neural networks for learning to detect retinal ves- vascular parameters for micro-CT imaging of vascular networks
sels in fundus images. arXiv:160304833v1 with multi-resolution. Med. Biol. Eng. Comput. 54(2–3), 511–
91. Li, Q., Feng, B., Xie, L., et  al.: A cross-modality learning 524 (2016)
approach for vessel segmentation in retinal images. IEEE Trans. 96. Piccinelli, M., Veneziani, A., Steinman, D.A., et al.: A framework
Med. Imaging 35(1), 109–118 (2016) for geometric analysis of vascular structures: application to cer-
92. Moeskops, P., Wolterink, J.M., van der Velden, B.H.M., et al.: ebral aneurysms. IEEE Trans. Med. Imaging. 28(8), 1141–1155
(2016) Deep Learning for Multi-task Medical Image Segmen- (2009)
tation in Multiple Modalities. In: Ourselin S, Joskowicz L, 97. Wu, X., Luboz, V., Krissian, K., et al.: Segmentation and recon-
Sabuncu MR, Unal G, Wells W (eds) Medical Image Computing struction of vascular structures for 3D real-time simulation. Med.
and Computer-Assisted Intervention—MICCAI 2016: 19th Inter- Image Anal. 15(1), 22–34 (2011)
national Conference, Athens, Greece, October 17–21, 2016. In: 98. Zhao, F., Sun, F., Hou, Y., et al.: (2017) A monocentric centerline
Proceedings, Part II. Springer International Publishing, Cham, extraction method for ring-like blood vessels. Med. Biol. Eng.
pp. 478–486. https://doi.org/10.1007/978-3-319-46723-8_55 Comput.
93. Annunziata, R., Trucco, E.: Accelerating convolutional sparse 99. Wong, W.C.K., So, R.W.K., Chung, A.C.S.: Principal curves for
coding for curvilinear structures segmentation by refining lumen center extraction and flow channel width estimation in
SCIRD-TS filter banks. IEEE Trans. Med. Imaging 35(11), 3-D Arterial networks: theory, algorithm, and validation. IEEE
2381–2392 (2016) Trans. Image Process. 21(4), 1847–1862 (2012)
94. Zhao, F., Liu, J., Qu, X., et al.: In vivo quantitative evaluation of 100. Zhang, J., Gao, Y., Park, S.H., et al.: Structured learning for 3D
vascular parameters for angiogenesis based on sparse principal perivascular spaces segmentation using vascular features. IEEE
component analysis and aggregated boosted trees. Phys. Med. Trans. Biomed. Eng. 64(12), 2803–2812 (2017)
Biol. 59(24), 7777–7791 (2014)

13

You might also like