Sie sind auf Seite 1von 4

MIXED DESIGN MIXDES 2011, 18th International Conference "Mixed Design of Integrated Circuits and Systems", June 16-18,

-18, 2011, Gliwice, Poland

Automatic Image Contrast Enhancement Method


for Liver Vasculature Detection
Marcin Rudzki
Silesian University of Technology
Dept. of Biomedical Engineering
Gliwice, Poland
Email: marcin.rudzki@polsl.pl

Abstract—This paper presents an automatic method for image It can also be used just for automatic window selection for
contrast enhancement necessary as a preprocessing step during medical images presentation.
vasculature detection in medical images. The method takes as in-
The paper is organized as follows: Section II briefly de-
put two volumetric images: one containing the original data from
a computed tomography (CT) study and the other one containing scribes the Hessian-based vessel enhancement filters found
segmented liver as a binary mask. The aim of the method is to in the literature, section III presents a new approach to
enhance the contrast in the original image based on blood vessels the automatic contrast enhancement. Testing and results are
intensities found in the organ defined by the provided mask. The presented in section IV. The paper ends with a short summary
methodology employs Frangi’s vesselness filter and the analysis
in section V.
of maximal intensity projection (MIP) images to calculate the
parameters of contrast enhancing function. The algorithm has
been evaluated on 40 CT datasets containing liver studies in II. R ELATED W ORK
the portal phase and is being tested as a preprocessing step for
automatic liver vasculature detection method being developed for In order to help the radiologists in their daily work sev-
a minimally-invasive liver surgery planning computer system. eral vessels detection and enhancement methods have been
Index Terms—contrast enhancement, Hessian based vesselness proposed [3], [4], [5]. They are based on the evaluation of
filter, liver vasculature detection, medical image processing.
Hessian eigenvalues and enhance bright cylindrical structures
within the image or use the Hessian eigenvalues evaluation
I. I NTRODUCTION as a part of processing [6]. Since only the eigenvalues of the
For the planning of liver tumor surgical resection it is Hessian are used in their response formula, such methods can
necessary to evaluate its vascularization (i.e. blood supply) detect all hyperintensive cylindrical structures independently
and its location within the liver segment or segments. Using of their intensity. This is not an issue while processing an-
Couinaud’s classification the liver can be divided into eight giographic images, where vessels are the only hyperintensive
independent segments. This classification is based on the lo- elongated objects. In case of the CT images other anatomical
cation of blood vessels inside the liver [1]. During a computed structures (like bones) exhibit similar density to contrast-
tomography (CT) study a patient is intravenously injected enhanced vessels, but usually have higher attenuation factor
with contrast agent which, by increasing the X-ray attenuation (image intensity) so may be erroneously treated as vessels.
factor, allows the blood vessels to be distinguished from the For this reason those methods require the robust masking of
liver tissue in the resulting image data due to increased voxel the region of interest (ROI) containing only the organ and its
intensities [2]. Depending on the acquisition protocol, time vessels or an intensity range rescaling function assuring that
and type of the contrast agent the vessels intensity may vary the intensities corresponding to vessels are the highest.
from one study to another.
The obtained image data is a series of 2D cross-sectional III. M ETHODOLOGY
images (slices) in which the pixel (voxel) value corresponds
The presented algorithm is provided with two volumetric
to the average X-ray attenuation value from a cuboid of fixed
(3D) images: I containing the data obtained from the CT study
physical dimensions and is expressed in Hounsfield Units
of a patient and M – a binary mask that defines the liver ROI.
(HU). The processing of those images requires the radiologist
The aims are:
to outline the anatomical structure of interest. However, the
manual segmentation of blood vessels is tedious and time 1) to find the intensity range for the image display (for
consuming mostly because of its tree-like structure cutting the example in a computer aided diagnosis station) in which
in-slice plane at various angles and in multiple areas. the lowest visible level corresponds to the organ tissue
The methodology presented in this paper is designed as and highest level to blood vessels,
a preprocessing step for a fully automatic liver vasculature 2) perform image intensity recalculation (windowing) that
detection method that is being developed for a computer would enhance all structures in the image I featuring
system aimed for minimally-invasive liver surgery planning. intensity level of vessels present within the ROI.

*QTv`B;?i Ü kyRR #v .2T`iK2Mi Q7 JB+`Q2H2+i`QMB+b  *QKTmi2` a+B2M+2- h2+?MB+H lMBp2`bBiv Q7 GQ/x j8


A. Vesselness filter
As a method for vessel detection Frangi’s vesselness filter
is used [4]. This section briefly describes its operation.
For each voxel located at point p = [x, y, z] a Hessian ma-
trix, containing second order partial derivatives, is calculated
⎡    ⎤
∂I  ∂I  ∂I 
∂x∂x ∂x∂y  ∂x∂z
⎢ p ⎥
⎢ ∂I  p
p
∂I  ∂I  ⎥

H(p) = ⎢ ⎢ ∂x∂y p ∂y∂y p ∂y∂z p ⎥
  (1)
⎣    ⎦
∂I  ∂I ∂I 
∂x∂z p ∂y∂z  ∂z∂z p
p
Fig. 1. Plot of generalized bell curves.
The image derivatives are computed by convolving the
image I with the second order derivative of a Gaussian kernel
with σ corresponding to the scale of image analysis. Eigenval- C. Calculation of window function parameters
ues of the Hessian are computed and sorted in ascending order
The window function has to be provided with three param-
by their absolute value so that |λ1 (p)| ≤ |λ2 (p)| ≤ |λ3 (p)|
eters: a – the exponent was fixed as constant with the value
holds true. According to their relations the information about
of 2, b – “halfwidth” – distance from the center to the 50%
the structure within the image can be obtained (Table I). The
decrease of curve value, c – center of the curve. The values
following denotations are used: L – small absolute value, H –
of parameters b and c have to be computed on the basis of the
a large absolute value, + a positive value, − a negative value.
input images.
The derivation of the relations can be found in [6].
In order to make the method insensitive to possible metallic
TABLE I implants and imperfections of the mask definition all structures
H ESSIAN EIGENVALUES AND CORRESPONDING IMAGE STRUCTURE [4] with the attenuation factor (image intensity) over 400HU are
removed. The removal procedure is as follows:
λ1 (p) λ2 (p) λ3 (p) structure
L L L noise
1) Generation of the binary mask MH that defines all
L L H- bright plane voxels with value ≥400HU:
L L H+ dark plane 
1 if I(p) >= 400
L H- H- bright cylinder MH (p) = (4)
L H+ H+ dark cylinder 0 otherwise
H- H- H- bright blob/ball
H+ H+ H+ dark blob/ball 2) Binary dilation with cubic structuring element S of size
3x3x3 voxels:
The vesselness measure is then voxel-wise calculated as MH ← MH ⊕ S (5)
follows: 3) Convolution of the mask MH with Gaussian kernel
having σ = 1.5mm:

0 if λ2 (p) ≥ 0 ∨ λ3 (p) ≥ 0

R2 (p)
R2 (p) MH ← MH ∗ G σ (6)
V (p) = A − B −
S 2 (p) (2)
⎩ 1 − e− 2α2 e 2β2 1 − e 2γ 2
4) Subtraction form the input image according to

where RA (p) = |λ2 (p)|


RB (p) = √ |λ1 (p)|
, S(p) = I(p) ← (I(p) − mI ) − (I(p) − mI ) · MH (p) + mI (7)
|λ3 (p)| , |λ2 (p)λ3 (p)|
||H(p)||, α = 0.5, and β = 0.5. Parameter γ influences the where mI is the minimal value of I within ROI defined
sensitivity of the filter to the magnitude of the second order by M .
image derivatives and can be automatically calculated as a half Next, in order to determine the value of parameter c the
of the maximal norm of all calculated Hessians [4]. following steps are performed:
1) Histogram ho (i) is calculated within the ROI of the
B. Window function
processed image I. Its maximum is found. This value
As an intensity selection (window function) and contrast determines the most common intensity value of the
enhancement a function organ’s tissue io .
1
y(x) =  2a (3) io = i : ho (i) = max(ho ) (8)
1 +  x−c  b
2) Background voxels in the processed image I are re-
defining the generalized bell curve has been chosen [7]. placed with the value io to minimize improper behavior
In Fig. 1 exemplary plots of the function are shown. of Frangi’s vesselness filter at the organ boundaries.
Parameters of the blue curve are a = 2, b = 25, c = 0 and of
green curve a = 2, b = 50, c = 100. I(p) ← io if M (p) = 0 (9)

je
3) Single iteration of Frangi’s vesselness filter is performed IV. T ESTING AND R ESULTS
on the resulting image with scale parameter of 2mm The implementation of the described method has been done
using the Insight Toolkit [8] medical image processing C++
V = FrangiFilter(I, α, β, γ, scale) (10) libraries and tested on 40 CT datasets containing liver studies
in the portal phase. The contrast between the vessels and
where: α = 0.5, β = 0.5, scale = 2mm and γ is set to
1 liver tissue varied from about 40HU to about 200HU. In-
4 to increase the filter sensitivity. plane resolution ranged from 0.7mm to 0.9mm and inter-slice
4) The response of the vesselness filter is thresholded at
spacing from 1.25mm to 7mm.
0.75 of its maximum and the resulting binary image
Exemplary input images (single slices) are shown in Fig-
defines vessels mask MV
ures 2, 4 and 6. The visible range is from -125HU to

1 if V (p) ≥ 0.75 · max(V ) 225HU. The resulting contrast enhanced images are presented
MV (p) = (11)
0 otherwise in Figures 3, 5 and 7, visible range is from 0 to 1.

5) Histogram hv of the processed image I inside regions


defined by the mask MV is calculated and the intensity
corresponding to its maximum is used as parameter c of
the window function

c = i : hv (i) = max(hv ). (12)

Alternatively, if the histogram hv has more than one


maximum the median of all image I intensities i within
the mask MV can be taken as the value of the parame-
ter c.
The calculation of parameter b is performed as follows:
Fig. 2. CT slice with high vessel contrast before processing.
1) Background removal using M (liver mask) and mI
(minimal image value within the liver ROI):

I(p) ← mI if M (p) = 0 (13)

2) From the resulting image a maximal intensity projection


(MIP) image is calculated:

MIPIX = max(I) (14)


x

and a corresponding MIP image defining 2D liver ROI


is calculated from the mask M :

MIPMX = max(M ) (15)


x

In the same manner the images MIPIY , MIPMY , MIPIZ Fig. 3. CT slice with high vessel contrast after processing, b = 90, c = 240
and MIPMZ are computed.
3) Histogram hX of MIPIX image is calculated only
within ROI defined by the corresponding mask MIPMX .
Intensity iX corresponding to the maximum of the
histogram hX is found in a similar manner as in (8).
The same operations are performed on MIPIY , MIPMY
and MIPIZ , MIPMZ images and intensity values iY and
iZ are calculated.
4) The value of parameter b of the window function is then

im = min(iX , iY , iZ ) (16)
b = c − im . (17)

The calculated parameters b and c are also employed as Fig. 4. CT slice with moderate vessel contrast before processing.
“window” and “level” parameters for the image display.

jd
in (19) and (18) were chosen experimentally to achieve best
results:
b ← 20 (18)
c ← im + 0.75 · b. (19)
Use of this “guessing” step allows the image contrast to be
correctly increased even if the CT image was improperly
acquired or the liver mask excludes most of the vasculature.
V. S UMMARY
An automatic method for contrast enhancement of blood
Fig. 5. CT slice with moderate vessel contrast after processing, b = 59, vessels in CT images was developed. The method uses two
c = 161
input images – CT series and a binary mask of organ of
interest. Searching for blood vessels inside the organ is done
using Frangi’s vesselness filter and the information about their
intensity is found. From MIP images of the organ volume the
information about intensity of the organ tissue is obtained.
Both serve as parameters of contrast-enhancing generalized
bell function. The method ignores all structures within ROI
whose intensity is higher than 400HU thus is insensitive to
presence of metallic implants and imperfections of the mask.
The algorithm has been tested on 40 CT datasets containing
liver studies in the portal phase and gives promising results.
It is being tested as a preprocessing step for an automatic
liver vessel detection method based on Frangi’s multiscale
vesselness filter. However further research is still required to
achieve robust operation of the automatic liver vasculature
detection system that is being developed for a minimally-
Fig. 6. CT slice with low vessel contrast before processing. Note presence
of metallic drain. invasive liver surgery planning computer system.
ACKNOWLEDGMENT
The work presented in this paper is a part of research grant
LIDER/03/47/L-1/NCBiR/2010 founded by the Polish Center
of Research and Development aiming at the development of
a computer system for liver surgery planning.
R EFERENCES
[1] C. Couinaud, “Liver anatomy: portal (and suprahepatic) or biliary seg-
mentation.” Digestive surgery, vol. 16, no. 6, pp. 459–467, 1999.
[2] M. Prokop, M. Galanski, A. J. Van Der Molen, and C. Schaefer-Prokop,
Spiral and Multislice Computed Tomograpgy of the Body. Thieme
Medical Publishers, 2003.
[3] Y. Sato, S. Nakajima, H. Atsumi, T. Koller, G. Gerig, S. Yoshida, and
R. Kikinis, “3D Multi-scale line filter for segmentation and visualization
of curvilinear structures in medical images,” in CVRMed-MRCAS ’97:
Proceedings of the First Joint Conference on Computer Vision, Virtual
Reality and Robotics in Medicine and Medial Robotics and Computer-
Fig. 7. CT slice with low vessel contrast after processing, b = 9, c = 108 Assisted Surgery. London, UK: Springer-Verlag, 1997, pp. 213–222.
replaced by “guessing” step to b = 20, c = 114 [4] A. F. Frangi, W. J. Niessen, K. L. Vincken, and M. A. Viergever,
“Multiscale vessel enhancement filtering,” in Lecture Notes in Computer
Science, Oct. 1998, vol. 1496, pp. 130+.
[5] R. Manniesing, M. Viergever, and W. Niessen, “Vessel enhancing diffu-
In seven cases of images with a very low vessel-liver sion. A scale space representation of vessel structures,” Medical Image
contrast (e.g. Fig. 6) the obtained values of b and c parameters Analysis, vol. 10, no. 6, pp. 815–825, Dec. 2006.
[6] K. Krissian, G. Malandain, N. Ayache, R. Vaillant, and Y. Trousset,
resulted in a very narrow intensity selection. This was caused “Model-Based Detection of Tubular Structures in 3D Images,” Computer
by a very low response of the Frangi’s vesselness filter. As Vision and Image Understanding, vol. 80, no. 2, pp. 130–171, Nov. 2000.
a remedy a “guessing” step have been added that is used [7] MathWorks. MATLAB fuzzy logic toolbox documentation. [Online].
Available: http://www.mathworks.com/help/toolbox/fuzzy/gbellmf.html
when the automatically calculated b < 20. Application of [8] Kitware. Insight Segmentation and Registration Toolkit. [Online].
such step is possible due to the fact, that the vessels are of Available: http://www.itk.org/
higher intensity than the surrounding tissue. The values present

j3

Das könnte Ihnen auch gefallen