Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_3592_Библиотеки_им_академика_М_И_Перельмана
.pdf
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 6.6. Comparison between ground-truth (green) and segmentation results (red): (a) s–t graph cut result,
(b) single-interface result and (c) double-interface result.
these methods, Zhang et al [28] proposed a region-based signed pressure force that
combines the Chan–Vese model and geodesic active contour model and utilises a
Gaussian fi lter to avoid re-initialisation of the signed distance function of the
generated level set; DRLSE is a typical edge-based method without the need for level
set re-initialisation using a distance regularisation term; VFC is a derived active
6-15
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 6.7. (a) Ground-truth. (b) Single-interface segmentation results. (c) Double-interface segmentation
results.
contour model with a proposed vector field convolution as a new external force; and
the star graph cut incorporates a star shape prior into the graph cut formulation. In
terms of graph-based techniques and shape prior integrations, the proposed method
is most similar to the star graph cut in principle. The proposed method is compared
to the star graph cut quantitatively and qualitatively with the full acquired data of
6-16
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Table 6.2. IVUS quantitative comparison. Mean value (standard deviation).
AMD HD AO Sens. Spec.
Texture-RBF method [17] 21.68 49.82 83.67 87.92 95.60
(13.37) (27.99) (8.92) 10.14 (4.70)
Single-interface with 1st GD 21.94 53.77 83.57 84.57 98.87
(19.36) (39.64) (13.37) (13.95) (1.55)
Proposed method 15.38 39.31 87.91 90.03 97.70
(9.92) (23.40) (6.94) (6.64) (5.15)
2283 images. The other methods used in the comparison are all initialisationdependent and thus we need to choose an appropriate iteration number for the best
results, in addition to a careful choice of an initialisation. With these considerations,
we randomly select 226 images from the total 2283 images to show their performance in dealing with the OCT segmentation.
In figure 6.9, a set of six OCT images shows the performance of both the star
graph cut [23] and the proposed method. The results for the star graph cut and the
proposed method are illustrated in columns (c) and (d), respectively, while the
original images and their ground-truth are listed in columns (a) and (b). It is
obviously apparent that most of the cases in the star graph cut are over-segmented
due to the use of ‘balloon’ force. In contrast, the proposed method generally
performs quite well, although the interference of various artefacts exists. However,
the proposed method is still affected in some situations due to the adversities in the
OCT modality. Several examples with inferior performance are presented in figure
6.10. In row 1, the results show that a serious bifurcation leads to poor performance
of the proposed method. In addition, residual blood (shown in row 2) and guide-wire
artefacts also cause undermining of border delineation, as shown in rows 3 and 4.
The star graph cut performs better than the proposed method in some of these cases.
The quantitative results are presented in table 6.3. The proposed method is
superior to the star graph cut in all metrics except for sensitivity. It is understandable
that the sensitivity of the star graph cut is a little greater than ours because the star
graph cut tends to over-segment.
To compare the proposed method to the deformable methods in capturing the
lumen area of OCT images, table 6.4 shows the quantitative results of three edge/
region-based methods along with the star graph cut and our method applied to 226
images. As they lack the help of a shape prior, the overall performance of these
methods is very poor. In contrast, the proposed method outperforms these
techniques significantly. In fact, the Chan–Vese method usually has the advantage
over the edge-based method because the region information can be extracted
appropriately. However, it is incapable of dealing with an OCT segmentation where
the induced artifact is very serious, so its AMD even reaches 39.99 and the AO is
merely 38.67%. Among these methods, VFC performs reasonably in terms of the use
of the vector field feature, whilst DRLSE cannot detect the lumen properly. In figure
6-17
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 6.8. Comparison between ground-truth (green) and segmentation results (red): (a) original image, (b)
texture-RBF method, (c) single-interface with 1st GD cost function and (d) proposed double-interface method.
6.11, four examples are presented to illustrate the relevant issues in these methods. In
column (c), due to the nature of the Chan–Vese method, the lumen area is always
segmented as the background while the bright area is detected. So, the method in [28]
is impotent in dealing with the OCT image. In contrast, VFC can perform well in
some cases but it meets great difficulties in a situation with serious artefacts, such as
6-18
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 6.9. Comparison with star graph cut. (a) Original image. (b) Ground-truth. (c) Star graph cut. (d)
Proposed method.
the last three cases. A quite similar performance occurs for the method of DRLSE.
However, its overall performance is poorer than VFC because it is sensitive to weak
edges. It is worth noting that DRLSE and VFC can work well when the artifact
interference is acceptable, such as the first case (figure 6.11).
6-19
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Figure 6.10. Cases with inferior performance of the proposed method. (a) Original image. (b) Ground-truth.
(c) Star graph cut. (d) Proposed method.
Table 6.3. Quantitative results with 2283 images for the star graph cut [23] and the proposed method.
Star graph cut [23] 5.85 22.11 91.27 96.52 98.07
Proposed method 4.94 19.09 92.51 96.03 98.85
AMD HD AO Sens. Spec.
(4.28) (17.91) (5.96) (4.59) (1.65)
(4.12) (20.13) (5.43) (4.66) (1.11)
6-20
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Table 6.4. Quantitative results with 226 images between the proposed method and the methods in [12, 13, 23]
and [28].
AMD HD AO Sens. Spec.
Improved Chan–Vese [28] 39.99 95.77 38.67 38.78 99.86
(20.54) (52.71) (30.54) (29.36) (0.81)
VFC [12] 14.34 50.03 80.05 85.65 97.67
(14.69) (41.53) (16.62) (17.72) (1.25)
DRLSE [13] 28.44 62.75 68.02 94.72 86.57
(9.73) (22.56) (10.83) (9.81) (3.65)
Star graph cut [23] 5.45 19.89 91.68 96.77 98.14
(3.34) (11.45) (5.39) (3.70) (1.03)
Proposed method 4.61 17.08 92.75 96.37 98.83
(2.49) (12.01) (4.45) (3.04) (1.03)
Figure 6.11. Results of the traditional methods in OCT: (a) original image, (b) ground-truth, (c) improved
Chan–Vese method, (d) DRLSE method, (e) VFC method and (f) proposed method.
6-21
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
6.7 Conclusion
Single- and double-interface segmentation methods for OCT and IVUS images were
presented. The segmentation problem is defined here as the delineation of the
media–adventitia border in IVUS and lumen–intima border in OCT. Images were
unravelled to polar coordinates, which facilitate the removing of the catheter ringdown artifact and converting the segmentation problem into finding the minimum
closed set graph that implies the border of interest. Steerable Gaussian derivative,
Gabor and local phase features are extracted from images that utilise image intensity
and texture information to highlight the desired border at different orientations and
scales. A new image feature is introduced that is derived from global interactions of
gradient vectors across the whole image domain. Laplacian diffusion is employed to
refine image features so as to produce more coherent segmentation.
For IVUS segmentation, an automatic double-interface segmentation method is
proposed, whose cost functions combine local and global image features and whose
geometric constraint is integrated in the graph construction. An auxiliary interface is
simultaneously searched to prevent undesirable image features from interfering with
the segmentation of the media–adventitia border. Qualitative and quantitative
comparison showed superior performance of the proposed method to the traditional
graph cut method or single-interface segmentation. For OCT segmentation, a single
-interface segmentation is proposed with a cost function defined at the zero-crossing
of the circulation density of the gradient convolution field. Experimental results in
OCT images demonstrate promising performance in comparison to the star graph
cut method. It is also evident that the proposed method takes great advantage of the
traditional edge/region-based deformable methods.
References
[1] Boykov Y and Funka-Lea G 2006 Graph cuts and efficient N-D image segmentation Int. J.
Comput. Vis.
[2] Boykov Y and Jolly M P 2001 Interactive graph cuts for optimal boundary and region
segmentation of objects in N-D images IEEE Int. Conf. on Computer Vision (Piscataway,
NJ: IEEE) pp 105–12
[3] Boykov Y and Kolmogorov V 2004 An experimental comparison of min-cut/max-fl ow
algorithms for energy minimization in vision IEEE Trans. Pattern Anal. Mach. Intell.
1124–37
[4] Chan T and Vese L 2001 Active contours without edges IEEE Trans. Image Process. 10
266–77
[5] Daugman J G 1985 Uncertainty relation for resolution in space, spatial frequency, and
orientation optimized by two-dimensional visual cortical filters J. Opt. Soc. Am. A
[6] Essa E, Xie X, Sazonov I and Nithiarasu P 2011 Automatic IVUS media–adventitia border
extraction using double interface graph cut segmentation IEEE Conf. on Image Processing
(Piscataway, NJ: IEEE) pp 69–72
[7] Essa E, Xie X, Sazonov I, Nithiarasu P and Smith D 2013 Shape prior model for media–
adventitia border segmentation in IVUS using graph cut Medical Computer Vision.
Recognition Techniques and Applications in Medical Imaging (Lecture Notes in Computer
70 109–31
2 1160–9
6-22
26
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
Science vol 7766) ed B Menze, G Langs, L Lu, A Montillo, Z Tu and A Criminisi (Berlin:
Springer) pp 114–23
[8] Fields D 1987 Relations between the statistics of natural images and the response properties
of cortical cells Opt. Soc. Am.
[9] Freeman W T and Adelson E H 1991 The design and use of steerable filters IEEE Trans.
Pattern Anal. Mach. Intell.
[10] Jones J L, Xie X and Essa E 2014 Combining region-based and imprecise boundary-based
cues for interactive medical image segmentation Int. J. Numer. Methods Biomed. Eng.
1649–66
[11] Kovesi P 1997 Symmetry and asymmetry from local phase Tenth Australian Joint Conf. on
Artificial Intelligence pp 185–90
[12] Li B and Acton S 2007 Active contour external force using vector field convolution for image
segmentation IEEE Trans. Image Process.
[13] Li C, Xu C, Gui C and Fox M 2010 Distance regularized level set evolution and its
application to image segmentation IEEE Trans. Image Process.
[14] Li K, Wu X, Chen D Z and Sonka M 2006 Optimal surface segmentation in volumetric
images—a graph-theoretic approach IEEE Trans. Pattern Anal. Mach. Intell.
[15] MacDonald D, Kabani N, Avis D and Evans A 2000 Automated 3-D extraction of inner
and outer surfaces of cerebral cortex from MRI NeuroImage
[16] Mulet-Parada M and Noble A 2000 2D + T acoustic boundary detection in echocardiog-
raphy Med. Image Anal.
[17] Papadogiorgaki M, Mezaris V, Chatzizisis Y S, Giannoglou G D and Kompatsiaris I 2008
Image analysis techniques for automated IVUS contour detection Ultrasound Med. Biol.
1482–98
[18] Sonka M, Zhang X, Siebes M, Bissing M S, DeJong S C, Collins S M and McKay C R 1995
Segmentation of intravascular ultrasound images: a knowledge-based approach IEEE Trans.
Med. Imaging
[19] Spreeuwers L and Breeuwer M 2003 Detection of left ventricular epi- and endocardial
borders using coupled active contours Int. Congr. Ser.
[20] Sun S, Sonka M and Beichel R 2013 Graph-based IVUS segmentation with efficient
computer-aided refinement IEEE Trans. Med. Imaging
[21] Takagi A, Hibi K, Zhang X, Teo T J, Bonneau H N, Yock P G and Fitzgerald P J 2000
Automated contour detection for high frequency intravascular ultrasound imaging: a
technique with blood noise reduction for edge enhancement Ultrasound Med. Biol.
1033–41
[22] Valois R L D, Albrecht D G and Thorell L G 1982 Spatial frequency selectivity of cells in
macaque visual cortex Vis. Res.
[23] Veksler O 2008 Star shape prior for graph-cut image segmentation Computer Vision—ECCV
2008 (Lecture Notes in Computer Science vol 5304) ed D Forsyth, P Torr and A Zisserman
(Berlin: Springer) pp 454–67
[24] Wahle A, Lopez J J, Olszewski M E, Vigmostad S C, Chandran K B, Rossen J D and Sonka
M 2006 Plaque development, vessel curvature, and wall shear stress in coronary arteries
assessed by x-ray angiography and intravascular ultrasound Med. Image Anal.
[25] Xie X and Mirmehdi M 2006 Magn etost atic field for the active contour model: a study
in convergence Proc. of the British Machine Vision Conf. (Surrey, UK: BMVA Press)
pp 14.1–10
14 719–32
4 2379–94
13 891–906
30
16 2096–106
19 3243–54
28 119–34
vol 12 340–56
4 21–30
1256 1147–52
32 1536–49
26
22 545–59
10 615–31
9
6-23
Vascular and Intravascular Imaging Trends, Analysis, and Challenges, Volume 1
https://t.me/medicina_free
[26] Xie X and Mirmehdi M 2008 MAC: magnetostatic active contour model IEEE Trans.
Pattern Anal. Mach. Intell.
[27] Zhang H, Essa E and Xie X 2013 Graph based segmentation with minimal user interaction
IEEE Conf. on Image Processing (Piscataway, NJ: IEEE) pp 4074–8
[28] Zhang K, Zhang L, Song H and Zhou W 2010 Active contours with selective local or global
segmentation: a new formulation and level set method Image Vis. Comput.
30 632–46
28 668–76
6-24
Соседние файлы в папке Библиотека им академика М.И. Перельмана