Огляд методів сегментації медичних зображень
This article provides an overview of the modern medical image segmentation methods. The most popular methods such as multi-atlas based methods and deep learning approach are considered in more details. In addition, this article overviews different steps of the multi-atlas based methods (MAS) in deta...
Збережено в:
| Дата: | 2018 |
|---|---|
| Автори: | , |
| Формат: | Стаття |
| Мова: | Англійська |
| Опубліковано: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2018
|
| Теми: | |
| Онлайн доступ: | https://journal.iasa.kpi.ua/article/view/126679 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Репозитарії
System research and information technologies| _version_ | 1866302170397147136 |
|---|---|
| author | Chapaliuk, Bohdan V. Zaychenko, Yuriy P. |
| author_facet | Chapaliuk, Bohdan V. Zaychenko, Yuriy P. |
| author_sort | Chapaliuk, Bohdan V. |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2018-04-12T11:42:34Z |
| description | This article provides an overview of the modern medical image segmentation methods. The most popular methods such as multi-atlas based methods and deep learning approach are considered in more details. In addition, this article overviews different steps of the multi-atlas based methods (MAS) in detail and shows which modern algorithms and approaches used in different steps of MAS to achieve state of the art results in the medical image segmentation task and how it affects the accuracy of the algorithm. Also, there is a brief description of the modern deep learning algorithms which are used for the medical image segmentation. Such type of algorithm is used as an independent algorithm or as a part of the MAS. Finally, this article summarizes described algorithms and evaluate which approaches promise to improve state of the art result of the medical image segmentation in the future. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2018.1.05 |
| first_indexed | 2025-07-17T10:23:28Z |
| format | Article |
| fulltext |
B.V. Chapaliuk, Yu.P. Zaychenko, 2018
72 ISSN 1681–6048 System Research & Information Technologies, 2018, № 1
UDC 683.519
DOI: 10.20535/SRIT.2308-8893.2018.1.05
MEDICAL IMAGE SEGMENTATION METHODS OVERVIEW
B.V. CHAPALIUK, Yu.P. ZAYCHENKO
Abstract. This article provides an overview of the modern medical image segmenta-
tion methods. The most popular methods such as multi-atlas based methods and
deep learning approach are considered in more details. In addition, this article over-
views different steps of the multi-atlas based methods (MAS) in detail and shows
which modern algorithms and approaches used in different steps of MAS to achieve
state of the art results in the medical image segmentation task and how it affects the
accuracy of the algorithm. Also, there is a brief description of the modern deep
learning algorithms which are used for the medical image segmentation. Such type
of algorithm is used as an independent algorithm or as a part of the MAS. Finally,
this article summarizes described algorithms and evaluates which approaches prom-
ise to improve state of the art result of the medical image segmentation in the future.
Keywords: medical image segmentation, multi-atlas based method, deep learning
approach
INTRODUCTION
In the last few years, there is huge progress in the computer vision task with the
help of the convolution neural networks (CNN) and an increasing size of the
labelled datasets for training. This success is attributed to the ability of the CNN
to learn a hierarchical representation of raw input data, without the usage of the
handcrafted features. Since the deep learning approaches become widely used in
computer vision field, there are lots of the works, which try to apply the same al-
gorithms and approaches for the medical image processing problem. Modern im-
age analysis technologies make good progress in the field of computer vision due
to deep learning algorithms and increasing number of available datasets as well.
One of the fundamental problems in the medical image processing is seg-
mentation. It refers to the process of tagging pixel region of interest (ROIs) with
biologically meaningful labels, for example, anatomical structures or tissue type.
As usual, to make modern deep learning algorithms work there should be a huge
amount of data which are manually labelled by the trained expert. Labelling and
segmenting dataset are expensive and complicated part of the application
of medical image analytics since medical image should be manually labelled by
the specialist in the medical area. Also, the labelled result may differ depending
on the operator, prone to error, not scalable and hard to reproduce. Furthermore,
the quality of dataset depends on the expert performance. Automatic [1] or semi-
automatic [2] segmentation algorithms can address these problems by speeding up
the process, reducing the cost and time, which should be spent by an expert, offer-
ing reliability, repeatability, and scalability.
There are several methods typically used for the medical image
segmentation: graph-based optimal image segmentation [3], multi-atlas based
methods [4], geometric deformable model approaches [5] and deep learning
Medical image segmentation methods overview
Системні дослідження та інформаційні технології, 2018, № 1 73
approaches [6]. This article considers the most popular modern approaches for
automatic image segmentation such as multi-atlas based methods and deep learn-
ing approach. Next chapter will review all the stages of the multi-atlas based
methods framework and will consider which approaches and algorithms are used
on each stage. The last chapter will briefly overview deep learning techniques,
which might be as independent segmentation algorithms or might be a part of the
multi-atlas based methods.
MULTI-ATLAS BASED METHODS
Multi-atlas based segmentation methods (MAS) are the class of methods which
aim to automatic segmentation of anatomical structure on the target image by
propagating a set of annotations from the set of atlas images to the new coordi-
nates on the novel image through image registration process. Typically, atlas im-
ages are manually segmented and annotated by the costly effort of the domain
expert who relies on the interactive visualisation software [7, 8].
There are several types of the atlas-based methods – probabilistic or
parametric methods which build a probabilistic representation of the set of the
atlases [9, 10] and nonparametric methods which use subset of atlases directly
(this type of the atlas-based methods are called multi-atlas based
methods) [11, 12].
There is a comparison between parametric and non-parametric methods [13]
which gives the evidence that the second type of methods shows better accuracy
than the first one. The success of these methods can be found in their possibility
to use the best-suited atlas images subset for segmenting each particular target
subject. Another explanation can be based on the fact that the registration of each
of the several best-suited atlases subsets to the target image is more robust than
the single image registration between the probabilistic atlas and the target image.
It is evident that the drawback of the atlas subset usage is an increasing require-
ment to the computation performance. However, it gives more accurate and robust
results. Therefore, the class of atlas-based methods should be chosen depending
on the application requirements. As far as the multi-atlas based methods show
better performance then probabilistic atlas-based method, this article will consider
them in more detail.
Formally, the goal of MAS is segmentation of some target image T using a
set of atlases mAA ,...,1 , and their corresponded maps label mLL ,...,1 which val-
ues can be 1 or 0 depend on whether voxel x belongs to the structure of interest
(1) or not (0). To achieve it, as usual MAS is involved into the next three steps:
atlas selection, image registration, label fusion step.
Additionally, modern algorithms can add some additional steps to the MAS
pipeline like online learning [14–17] and post-processing steps [18–21].
Atlas selection step
In the scope of this step, MAS algorithms select the subset of the atlas images that
are the most anatomically similar to the target subject TS in ),...,1( m . All avail-
able atlas images are not used in MAS for the several reasons: first, this approach
will improve computation efficiency of the algorithm which might be very impor-
tant for the application which has strict time constraints. Second, excluding the
B.V. Chapaliuk, Yu.P. Zaychenko
ISSN 1681–6048 System Research & Information Technologies, 2018, № 1 74
irrelevant images for the target subject ST would improve final algorithm accu-
racy.
At the early time in the scope of the atlas selection step the subset of the at-
las images are picked up randomly [11]. Under random selection of the atlas
images, the accuracy of the algorithm improves with increasing number of the
atlases still the number of atlases should not be too high to avoid accuracy
worsening by introduced not well-suited candidates. It occurs due to the
combining result of the multiple atlases leads to correct the errors of the one
individual atlas. This idea is fundamental for the multi-atlas based algorithms.
The MAS algorithm segmentation performance can be improved in
comparison with the randomly chosen atlases if the algorithm chooses the set of
atlases which best-representing anatomy of the target image. That is because the
non-related anatomical characteristic will be filtered out. The main issue, which
the atlas selection step is solving, consists of defining a function that reflects the
similarity between an atlas iA and the target image .T To resolve this task there
are lots of the metrics used. The metrics might inlclude the similarity measures on
the image intensities or normalized mutual information [22], registration
consistancy [23], cross-correlation of intensity values [24].
More recently, there are several works which defined another approach to
increase performance of atlas selection step by using clustering [25]. In this paper,
the authors propose to define the vector of a pair atlas agreement factor between
two atlases and apply the k-means algorithm to this feature vector. Furthermore,
the ranking parameter, which evaluates the mean SSD between cluster means, is
aplied to the result clusters. Such approach improves the total performance of
their MAS algorithm.
On the other hand, atlas selection might be treated as a learning algorithm.
Such approach utilises the necessity of the manual segmentation of the atlas
image which reduces the cost and effort of the MAS algorithm development and
also tends to be more computation efficiently than discussed approach, especially
on the large dataset. For example, this paper [26] introduced Neighborhood
Approximation Forest, a supervised learning algorithm which is inspired by the
random forest tree algorithm [27]. This algorithm can incorporate distances that
are defined on semantic information and relate them to the space of appearance-
based feature. Another algorithm, which uses the same metrics for evaluating
similarity between target image T and atlas image, is SVM which show state of
the art result for this step [28].
To sum up, atlas selection step is an important step of MAS which can affect
the final algorithm accuracy. To achieve higher accuracy on the segmentation for
atlas selection can be used different metrics. However, the most promised
approach is the learning algorithms which use distance metrics like SSD. On the
other hand, an approach of atlas selection algorithm might vary depending on the
image registration algorithm is used and specific application requirement like
computation time of segmentation.
Image registration step
By the registration step in the scope of the MAS is meant the task of establishing
a spatial correspondence between the target image and set of atlases. As a result,
the labels can be directly propagated. It means that the possible values of labels
on each target pixel correspond to the pixel on the target image.
Medical image segmentation methods overview
Системні дослідження та інформаційні технології, 2018, № 1 75
More formally [4], set of atlases A correspondences with the target image
T can be described as follows:
)})(reg)(,(sim{max
ii ATL ,
where operation is some similarity measurement term between two images,
operation reg is a regularization term, which can control the flexibility of the
transformation with constant , iA — atlas image, is a spatial transformation.
The result of the label map calculation will be used on the label fusion step
according to the computed transformation. More formally, it will be used in form
)(
~
il LL . An extensive review of the existing spatial transformation, similarity
function and regularization terms can be found in [29]. An extensive comparison
review of the image registration methods on the publicly available datasets can be
found in [30].
There are two types of the registration types according to its transformation
model: linear and deformable registration [29].
Typically, in MAS one single registration is computed between each atlas
and the target image. The usage of the atlas sets tends to improve the performance
of the registration algorithm and give more robust result as was discussed at the
beginning of the paper.
In conclusion, the registration algorithms map the selected atlases to the tar-
get image and create a label map with correspondence between pixels on the atlas
images and target image T. Recently, the most popular and accurate methods of
image registration are methods which are based on the deformable model.
Label fusion step
Label fusion is an important part of the MAS algorithm pipeline step. During this
step, the final segmentation on the target image is produced. It uses the result
mapping between each atlas and the target image T which is got from the image
registration step. More formally [4], each atlas iA
~
and the label map iL
~
are regis-
tered to the target image T after image registration step. The segmentation result
on each target voxel x of the target image T evaluated by combining labels on
the corresponding voxel location )(xLi in the atlas image. The earliest and sim-
plest algorithm of the label fusion are best atlas selection [31] and majority voting
rule [11, 32]. Specifically, it can be described by the next formula:
TSi
i
l
lxLxF ))((maxarg
}1,0{
,
where is the function which returns 1 if the argument is true and 0 otherwise,
and TS is the subset of the selected atlases indexes.
Even though the majority voting rule is a simple one, it gives better accuracy
than any other method which uses a single atlas for image registration
[31]. Nevertheless, in the real-world application some correlation in the error pat-
tern might exist, and therefore, more robust label fusion algorithm is needed to
compensate it.
B.V. Chapaliuk, Yu.P. Zaychenko
ISSN 1681–6048 System Research & Information Technologies, 2018, № 1 76
According to the paper [4] there are three categories of the label fusion
methods used in the modern MAS applications: weighted voting approaches,
probabilistic approaches, machine-learning based approaches.
Weighted voting approaches. These approaches use weights for evaluate
importance of each atlas and select the best one. Weights reflect the similarity
between the atlas and the target image and might be global or local. Conse-
quently, there are two types of methods: methods which compute weights based
on each atlas independently and the methods that compute weights to minimise
correlation between error patterns of atlas pairs [33, 34].
In case of the independent weighting strategy, each atlas is evaluated indi-
vidually with respect to the weights. Weights represent the local importance of
atlas and computed based on the similarity function between the local image
patches. Typically, as similarity measurement metrics used cross-correlation, mu-
tual information metrics or the sum of squared differences [33], or an empirical
measurement might be used [35]. For example, similarity measurement metrics
between local image patches of the target image and atlas can be negative SSD
exponent [4]:
XNy
ii yAYTx 2))()((exp)( ,
where is a normalization parameter, xN is the spatial neighborhood which de-
fines the image patches centered at x .
As a result, the target label will be computed as follows:
TSi
ii
l
lxLxxF ))(()(maxarg)(
,
where )(xi is weight which denotes i -th atlas importance in evaluating the tar-
get label at the region location x , is the function which returns 1 if the argu-
ment is true and 0 otherwise, and TS is the subset of the selected atlases indexes.
There might be two strategies to denote correspondence between the target
image and atlas image: one-to-one correspondence and one-to-many correspon-
dences. The one-to-many correspondence is considered to be better than one-to-
one correspondences as far as it has better robustness because of considering the
atlas labels in the spatial neighbourhood [36].
For the one-to-many correspondence strategy [4], the target label calculation
will look as follows:
t xSi Ny
ii lyLyxxF ))((),(maxarg)(
,
where ),( yxi is weight which denotes estimated segmentation performance for
the operation of assigning the atlas label at the point xNy to the target label at
the point x , xN is spatial neighborhood for potential atlas correspondences
search, is the function which returns 1 if the argument is true and 0 otherwise,
and TS is the subset of the selected atlases indexes.
Medical image segmentation methods overview
Системні дослідження та інформаційні технології, 2018, № 1 77
The independent weighting strategy works well when the anatomical image
characteristics are equally distributed among the atlases, however, in the real-
world usually this assumption does not work, and some anatomical characteristics
and features may be overrepresented in the set of atlases. To deal with this prob-
lem the joint weighting strategy is used. [37] The main idea behind weighting
strategy is to minimise the correlation of participating atlases during the weights
choosing step instead of computing them independently. This approach enforces
that the most represented anatomical characteristic in the dataset not accumulated
in most of the calculated weights.
Probabilistic approaches. These approaches use Bayes` probability rules to
select the best one label. There are two core algorithms family: STAPLE
algorithm [38, 39] that directly estimates parameters performance that best suit to
the probabilistic estimate of the target labels, and generative probabilistic model
[40] which considers weighted voting rule from a Bayesian perspective.
Machine-learning based approaches. These approaches use supervised
learning to evaluate the relationship between the appearance feature and the ana-
tomical feature. [41]. To achieve optimal performance the set of atlases with their
corresponding label are used to lean the classification rules. In comparison to the
traditional MAS label fusion algorithm, which gets the target labels from the im-
age registration step, machine learning approaches can capture more complex re-
lationships between image and labels. The machine learning approaches will be
considered in more detail in the next chapter, where the deep learning approaches
will be discussed.
DEEP LEARNING APPROACH
One of the limitations of the previously discussed methods is the inability of
adapting themselves to the data at hand. This means that the power of feature rep-
resentation might vary across the different kind of image data. The handcrafted
feature and representations depend on the expert performance, which may vary
between different domain experts. Furthermore, manually feature creation cannot
tend to the creation of a complex feature pattern. So far as these problems can be
resolved by a deep learning approaches, it gains popularity in the research com-
munity. Recently, deep learning has become a hot topic in machine learning [42],
computer vision and biomedical image processing [43]. Likewise, deep learning
approach with its ability to learn hierarchy from raw data tends to improve the
overall performance of the task [44] meanwhile reduce the cost of application de-
velopment since the manual feature design is not required anymore. In addition,
there are the deep learning approaches which outperform the classical methods
like MAS with a large margin [45].
The work [43] shows an attempt to use the deep learning approach for the
multiple organ detections using the 4D patient data. The authors used an unsuper-
vised technique to create a feature representation for their data, and after that use
them in the probabilistic patch-based methods. They use stacked sparse autoen-
coder (SSAE) to extract the feature from the dataset.
There is also another work, which extends such approach and uses the stack
autoencoders as a part of the MAS and deformable model methods [46]. The au-
thors perform the supervised fine-tuning which is adapted by stacking another
B.V. Chapaliuk, Yu.P. Zaychenko
ISSN 1681–6048 System Research & Information Technologies, 2018, № 1 78
classification output layer on the top of the encoding part of the SSAE. The result
of the work shows that the SSAE outperform other classical methods including
the MAS and deformable model methods with handcrafted feature representation.
Another approach for applying deep learning algorithms for the medical im-
age segmentation is to use the power of convolutional neural networks (CNN). In
[47] it was proposed the U-Net architecture for the CNN which is an extension of
the fully convolution neural network [48]. The authors report that they achieve
state of the art results on the two challenging light data sets and tasks: segmenta-
tion of neuronal structure in EM stacks and cell segmentation in light microscopy
challenge from ISBI cell tracking challenge. Consequently, this CNN has a good
performance even in the small datasets.
One of the disadvantages of the U-Net CNN is an ability to work only with
2D data while most of the clinical data consist of the 3D volumes. To make
the CNN to process the 3D image the V-Net CNN architecture was introduced [49].
This architecture achieves the superior result on the PROMISE12 dataset [50].
After all, the deep learning approaches show the state of the art result and
promise to achieve state of the art performance on the most biomedical processing
task. These approaches might be used as a part of the classical segmentation ap-
proaches, like MAS as well as independently.
CONCLUSION
This article considers auto-segmentation methods which is one of the most widely
used methods in the biomedical image processing. For the last 10 years the multi-
atlas based methods have been rapidly developed and achieved great performance
on the different type of medical data. MAS methods consist of the three main
steps: label selection, image registration and label fusion steps. The label selection
step selects the most anatomically similar atlases to the target image. It has been
shown, that the best approach is to select a subset of the anatomically similar at-
lases rather than use all available atlases. The registration step builds the map of
correspondence between each atlas and the target image. This step has a huge im-
pact on the overall MAS algorithm performance, so that the image registration
approach should be chosen carefully according to the medical image processing
performance. The label fusion step combines the individual decisions with the
multiple atlas to decide which label should be applied to the certain point in the
image. The most widely used algorithms for the label fusion are vote weighting,
probabilistic weighted vote, based on the Bayes` framework. There are algorithms
which use machine learning approach also.
On the other hand, recently the deep learning approach has been largely used
in the biomedical image processing field and shown much promising result. There
are several CNNs architectures which have achieved state of the art results on the
different challenge dataset. Recent development shows that different supervised
machine learning and unsupervised learning techniques tend to be used in the dif-
ferent steps of the MAS algorithm to improve the overall accuracy.
REFERENCES
1. Wu Z. Segmenting hippocampal subfields from 3T MRI with multi-modality images
/ Z. Wu, Y. Gao, F. Shi et al. // Segmenting hippocampal subfields from 3T MRI
with multi-modality images. — Vol. 43. — 2018. — P.10–22.
Medical image segmentation methods overview
Системні дослідження та інформаційні технології, 2018, № 1 79
2. You W. Semi-automatic segmentation of the placenta into fetal and maternal
compartments using intravoxel incoherent motion MRI / W. You, N. Andescavage,
Z. Zun, C. Limperopoulos // Medical Imaging. — Vol. 10137. — 2017.
3. Oguz I. LOGISMOS: A Family of Graph-Based Optimal Image segmentation
methods / I. Oguz, H. Bogunovic, S. Kashyap et al. // Medical Image
Recognition, Segmentation and Parsing Machine Learning and Multiple Object
Approaches. — Academic Press, 2016. — P.179–206.
4. Sanroma G. Multiple-atlas segmentation in medical imaging / G. Sanroma, G. Wu,
M. Kim, M.A. González et al. // Medical Image Recognition, Segmentation and
Parsing. — Academic Press, 2016. — P.231–257.
5. Carass A. An Overview of the Multi-Object Geometric Deformable Model
Approach in Biomedical Imaging / A. Carass, J.L. Prince // Medical Image
Recognition, Segmentation and Parsing Machine Learning and Multiple Object
Approaches. — Academic Press, 2016. — P.259–279.
6. Zhou K. Deep Learning for Medical Image Analysis / K. Zhou, H. Greenspan,
D. Shen. — Academic Press, 2017.
7. Criminisi A. Geos: Geodesic image segmentation / A. Criminisi, T. Sharp, A. Blake
// Computer Vision — ECCV 2008. ECCV 2008. Lecture Notes in Computer
Science. — Springer, 2008. — P.99–112.
8. Heiberg E. Design and validation of segment-freely available software for
cardiovascular image analysis / E. Heiberg, J. Sjogren, M. Ugander et al. // BMC
medical imaging. — Vol. 10. — 2010.
9. Pohl K. Bayesian model for joint segmentation and registration / K. Pohl, J. Fisher,
L. Grimson et al. // Neuroimage. — Vol. 31. — 2006. — P. 228–239.
10. Yeo B. Effects of registration regularization and atlas sharpness on segmentation
accuracy / B. Yeo, M. Sabuncu, R. Desikan et al. // Medical image analysis. —
Vol. 12. — 2008. — P. 603–615.
11. Heckemann R. Automatic anatomical brain MRI segmentation combining label
propagation and decision fusion / R. Heckemann, J. Hajnal, P. Aljabar et al. //
Neuroimage. — N. 33. — 2006. — P. 115–126.
12. Heiberg E. Design and validation of Segment – freely available software for
cardiovascular image analysis / E. Heiberg, J. Sjögren, M. Ugander, M. Carlsson
et al. // BMC Medical Imaging. — Vol. 10. — 2010. — P. 1.
13. Babalola K. An evaluation of four automatic methods of segmenting the subcortical
structures in the brain / K. Babalola, B. Patenaude, P. Aljabar et al. //
NeuroImage. — Vol. 47, N. 4. — 2009. — P. 1435–1447.
14. Langerak T. Label Fusion in Atlas-Based Segmentation Using a Selective and
Iterative Method for Performance Level Estimation (SIMPLE) / T. Langerak,
U. van der Heide, A. Kotte et al. // IEEE Transactions on Medical Imaging. —
Vol. 29. — 2010. — P. 2000–2008.
15. Hao Y. Local label learning (LLL) for subcortical structure segmentation:
Application to hippocampus segmentation / Y. Hao, T. Wang, X. Zhang et al. //
Human Brain Mapping. — Vol. 35. — 2014. — P. 2674–2697.
16. Wolz R. LEAP: Learning embeddings for atlas propagation / R. Wolz, P. Aljabar,
J. V. Hajnal et al., the Alzheimer's Disease Neuroimaging Initiative //
NeuroImage. — Vol. 49, N. 2. — 2010. — P. 1316–1325.
17. Chakravarty M. Performing label-fusionbased segmentation using multiple
automatically generated templates / M. Chakravarty, P. Steadman, M. Eede et al.
// Human brain mapping. — Vol. 34. — 2013. — P. 2635–2654.
18. Yushkevich P. Nearly automatic segmentation of hippocampal subfields in in vivo
focal T2-weighted MRI / P. Yushkevich, H. Wang, J. Pluta et al. // Neuroimage.
— Vol. 53. — 2010. — P. 1208–1224.
B.V. Chapaliuk, Yu.P. Zaychenko
ISSN 1681–6048 System Research & Information Technologies, 2018, № 1 80
19. Nouranian S. A multi-atlas-based segmentation framework for / S. Nouranian,
S. Mahdavi, I. Spadinger et al. // IEEE Transactions on Medical Imaging. —
Vol. 34. — 2015. — P. 950–961.
20. Wang L. Segmentation of neonatal brain MR images using patch-driven level sets /
L. Wang, F. Shi, G. Li et al. // Neuroimage. — Vol. 84. — 2014. — P. 141–158.
21. Asman A. Out-of-atlas likelihood estimation using multi-atlas segmentation /
A. Asman, L. Chambless, R. Thompson, B. Landman // Medical physics. —
Vol. 40, N. 4. — 2013.
22. Studholme C. An overlap invariant entropy measure of 3D medical image alignment
/ C. Studholme, D.L. Hill, D. Hawkes // Pattern Recognition. — Vol. 32. —
1999. — P. 71–86.
23. Heckemann R. The mirror method of assessing segmentation quality in atlas label /
R. Heckemann, A. Hammers, P. Aljabar et al. // Biomedical Imaging: From Nano
to Macro, 2009. ISBI '09. IEEE International Symposium on. — 2009. —
P. 1194–1197.
24. Avants B. The optimal template effect in hippocampus studies of diseased
populations / B. Avants, P. Yushkevich, J. Pluta et al. // NeuroImage. — Vol. 49,
N. 3. — 2010. — P. 2457–2466.
25. Nouranian S. A Multi-Atlas-Based Segmentation Framework for Prostate
Brachytherapy / S. Nouranian, S.S. Mahdavi, I. Spadinger et al. // IEEE
Transactions on Medical Imaging. — Vol. 34, N. 4. — 2015. — P. 950–961.
26. Konukoglu E. Neighbourhood approximation using randomized forests /
E. Konukoglu, B. Glocker, D. Zikic, A. Criminisi // Medical Image Analysis. —
Vol. 17, N. 7. — 2013. — P. 790–804.
27. Liaw A. Classification and Regression by randomForest / A. Liaw, M. Wiener //
R news. — Vol. 2, N. 3. — 2002. — P. 18–22.
28. Sanroma G. Learning to rank atlases for multiple-atlas segmentation / G. Sanroma,
G. Wu, Y. Gao, D. Shen // IEEE Trans Med Imaging. — Vol. 33. — 2014. —
P. 1939–1953.
29. Sotiras A. Deformable Medical Image Registration: A Survey / A. Sotiras,
D. Christos, P. Nikos // IEEE Transactions on Medical Imaging. — Vol. 32. —
2013. — P. 1153–1190.
30. Ou Y. Comparative Evaluation of Registration Algorithms in Different Brain
Databases With Varying Difficulty: Results and Insights / Y. Ou, H. Akbari,
M. Bilello et al. // IEEE Transactions on Medical Imaging. — Vol. 33. — 2014.
— P. 2039–2065.
31. Rohlfing T. Evaluation of atlas selection strategies for atlas-based image
segmentation with application to confocal microscopy images of bee brains /
T. Rohlfing, R. Brandt, R. Menzel, Jr. CR Maurer // Neuroimage. — Vol. 21. —
2004. — P. 1428–1442.
32. Klein A. Automated brain labeling with multiple atlases / A. Klein, B. Mensh,
S. Ghosh et al. // BMC Medical Imaging. — Vol. 5. — 2005. — P. 7.
33. Artaechevarria X. Efficient classifier generation and weighted voting for atlas-based
segmentation: Two small steps faster and closer to the combination oracle /
X. Artaechevarria, A. Muñoz-Barrutia, C. Ortiz-de Solorzano // Medical Imaging
2008: Image Processing. —2008.
34. Wan J. Automated reliable labeling of the cortical surface / J. Wan, A. Carass,
S. Resnick, J. Prince // Biomedical Imaging: From Nano to Macro. — 2008. —
P. 440–443.
35. Zhang D. Confidence-guided sequential label fusion for multi-atlas based / D. Zhang,
G. Wu, H. Jia, D. Shen // MICCAI. — Vol. 6893. — 2011. — P. 643–650.
36. Rousseau F. A supervised patch-based approach for human brain labeling /
F. Rousseau, P.A. Habas, C. Studholme // IEEE Transactions on Medical
Imaging. — Vol. 30, N. 10. — 2011. — P. 1852–1862.
Medical image segmentation methods overview
Системні дослідження та інформаційні технології, 2018, № 1 81
37. Wang H. Regression-based label fusion for multi-atlas segmentation / H. Wang,
J. Suh, S. Das et al. // Computer Vision and Pattern Recognition (CVPR), 2011
IEEE Conference on. — 2011. — P. 1113–1120.
38. Warfield S.K. Simultaneous truth and performance level estimation (STAPLE): an
algorithm for the validation of image segmentation / S.K. Warfield, K.H. Zou,
W.M. Wells // IEEE Transactions on Medical Imaging. — Vol. 23. — 2004. —
P. 903–921.
39. Asman A.J. Hierarchical performance estimation in the statistical label fusion
framework / A.J. Asman, B.A. Landman // Medical Image Analysis. — Vol. 18,
N. 7. — 2014. — P. 1070–1081.
40. Sabuncu M. A Generative Model for Image Segmentation Based on Label Fusion /
M. Sabuncu, Y. Thomas, K. Leemput et al. // IEEE Transactions on Medical
Imaging. — Vol. 29. — 2010. — P. 1714–1729.
41. Wang H. Multi-atlas segmentation with learning-based label fusion / H. Wang,
Y. Cao, T. Syeda-Mahmood // Machine Learning in Medical Imaging. —
2014. — P. 256–263.
42. Bengio Y. Representation Learning: A Review and New Perspectives / Y. Bengio,
A. Courville, P. Vincent // IEEE Transactions on Pattern Analysis and Machine
Intelligence. — Vol. 35, N. 8. — 2013. — P. 1798–1828.
43. Shin Hoo-Chang Stacked Autoencoders for Unsupervised Feature Learning and
Multiple Organ Detection in a Pilot Study Using 4D Patient Data / Hoo-Chang
Shin, Matthew R. Orton, David J. Collins, Simon J. Doran, Martin O. Leach //
IEEE Transactions on Pattern Analysis and Machine Intelligence. — Vol. 35,
N. 8. — 2013. — P. 1930–1943.
44. Krizhevsky A. ImageNet Classification with Deep Convolutional Neural Networks /
A. Krizhevsky, I. Sutskever, G. Hinton // Advances in Neural Information
Processing Systems 25. — Curran Associates, Inc., 2012. — P.1097–1105.
45. Ciresan D. Deep Neural Networks Segment Neuronal Membranes in Electron
Microscopy Images / D. Ciresan, A. Giusti, L. Gambardella, J. Schmidhuber //
Advances in Neural Information Processing Systems 25. — Curran Associates,
Inc., 2012. — P.2843–2851.
46. Guo Y. Deformable MR Prostate Segmentation via Deep Feature Learning and
Sparse Patch Matching / Y. Guo, Y. Gao, D. Shen // Deep Learning for Medical
Image Analysis. — Academic Press, 2017. — P.221–247.
47. Ronneberger O. U-Net: Convolutional Networks for Biomedical Image
Segmentation / O. Ronneberger, P. Fischer, T. Brox // Medical Image Computing
and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in
Computer Science. — Springer, 2015. — P.234–241.
48. Jia Y. Caffe: Convolutional architecture for fast feature embedding / Y. Jia,
E. Shelhamer, J. Donahue et al. // MM '14 Proceedings of the 22nd ACM
international conference on Multimedia. — 2014. — P. 675–678.
49. Milletari F. V-Net: Fully Convolutional Neural Networks for Volumetric Medical
Image Segmentation / F. Milletari, N. Navab, Sed-Ahmad Ahmadi // 3D Vision
(3DV), 2016 Fourth International Conference on. — 2016. — P. 565–571.
50. Litjens G. Evaluation of prostate segmentation algorithms for MRI: the PROMISE12
challenge / G. Litjens, R. Toth, W. van de Ven et al. // Medical image analysis.
— Vol. 18. — 2014. — P. 1361–8423.
Received 01.11.2017
|
| id | journaliasakpiua-article-126679 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:23:28Z |
| publishDate | 2018 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/40/d00c0099c2409266b137716f57719740.pdf |
| spelling | journaliasakpiua-article-1266792018-04-12T11:42:34Z Medical image segmentation methods overview Обзор методов сегментации медицинских изображений Огляд методів сегментації медичних зображень Chapaliuk, Bohdan V. Zaychenko, Yuriy P. medical image segmentation multi-atlas based method deep learning approach сегментация медицинских изображений много-атласные методы глубинное обучение сегментація медичних зображень багато-атласні методи глибинне навчання This article provides an overview of the modern medical image segmentation methods. The most popular methods such as multi-atlas based methods and deep learning approach are considered in more details. In addition, this article overviews different steps of the multi-atlas based methods (MAS) in detail and shows which modern algorithms and approaches used in different steps of MAS to achieve state of the art results in the medical image segmentation task and how it affects the accuracy of the algorithm. Also, there is a brief description of the modern deep learning algorithms which are used for the medical image segmentation. Such type of algorithm is used as an independent algorithm or as a part of the MAS. Finally, this article summarizes described algorithms and evaluate which approaches promise to improve state of the art result of the medical image segmentation in the future. Приведен обзор современных методов сегментации медицинских изображений, наиболее популярные методы, такие как многоатласные методы (МАМ) и методы сегментации на базе глубинного обучения. Подробно изложен каждый из шагов МАМ, алгоритмы и подходы, что используются для достижения наибольшей точности сегментации, а также влияние выбора каждого алгоритма определенного шага МАМ на суммарную точность работы алгоритма. Показаны современные подходы глубинного обучения, применяемые для сегментации медицинских изображений. Такие алгоритмы обычно используются как самостоятельные независимые алгоритмы, но могут использоваться и как часть МАМ. Оценены подходы, которые могут помочь повысить точность сегментации медицинских изображений в будущем. Наведено огляд сучасних методів сегментації медичних зображень. Розглянуто найбільш популярні методи, такі як багатоатласні методи (БАМ) та методи сегментації на базі глибинного навчання. Детально викладено кожен із кроків БАМ, алгоритми та підходи, які використовуються для досягнення найбільшої точності сегментації, а також вплив вибору кожного алгоритму певного кроку БАМ на сумарну точність роботи алгоритму. Показано сучасні підходи глибинного навчання, що застосовуються для сегментації медичних зображень. Такі алгоритми зазвичай використовуються як самостійні незалежні алгоритми, але можуть використовуватися і як частина БАМ. Оцінено підходи, які можуть допомогти підвищити точність сегментації медичних зображень в майбутньому. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2018-03-20 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/126679 10.20535/SRIT.2308-8893.2018.1.05 System research and information technologies; No. 1 (2018); 72-81 Системные исследования и информационные технологии; № 1 (2018); 72-81 Системні дослідження та інформаційні технології; № 1 (2018); 72-81 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/126679/123509 Copyright (c) 2021 System research and information technologies |
| spellingShingle | сегментація медичних зображень багато-атласні методи глибинне навчання Chapaliuk, Bohdan V. Zaychenko, Yuriy P. Огляд методів сегментації медичних зображень |
| title | Огляд методів сегментації медичних зображень |
| title_alt | Medical image segmentation methods overview Обзор методов сегментации медицинских изображений |
| title_full | Огляд методів сегментації медичних зображень |
| title_fullStr | Огляд методів сегментації медичних зображень |
| title_full_unstemmed | Огляд методів сегментації медичних зображень |
| title_short | Огляд методів сегментації медичних зображень |
| title_sort | огляд методів сегментації медичних зображень |
| topic | сегментація медичних зображень багато-атласні методи глибинне навчання |
| topic_facet | medical image segmentation multi-atlas based method deep learning approach сегментация медицинских изображений много-атласные методы глубинное обучение сегментація медичних зображень багато-атласні методи глибинне навчання |
| url | https://journal.iasa.kpi.ua/article/view/126679 |
| work_keys_str_mv | AT chapaliukbohdanv medicalimagesegmentationmethodsoverview AT zaychenkoyuriyp medicalimagesegmentationmethodsoverview AT chapaliukbohdanv obzormetodovsegmentaciimedicinskihizobraženij AT zaychenkoyuriyp obzormetodovsegmentaciimedicinskihizobraženij AT chapaliukbohdanv oglâdmetodívsegmentacíímedičnihzobraženʹ AT zaychenkoyuriyp oglâdmetodívsegmentacíímedičnihzobraženʹ |