HAL Id: inria-00615108
https://hal.inria.fr/inria-00615108
Submitted on 17 Aug 2011
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Fully automatic anatomical, pathological, and functional
segmentation from CT scans for hepatic surgery
Luc Soler, Hervé Delingette, Grégoire Malandain, Johan Montagnat, Nicholas
Ayache, Christophe Koehl, Olivier Dourthe, B. Malassagne, M. Smith, Didier
Mutter, et al.
To cite this version:
Luc Soler, Hervé Delingette, Grégoire Malandain, Johan Montagnat, Nicholas Ayache, et al.. Fully
automatic anatomical, pathological, and functional segmentation from CT scans for hepatic surgery.
Computer Aided Surgery, Taylor & Francis, 2001, 6 (3), pp.131-42. �inria-00615108�
Fully automatic anatomical, pathological,
and functional segmentation from CT scans for hepatic surgery
Luc Soler
a
(PhD), Herve Delingette
b
(PhD), Gregoire Malandain
b
(PhD), Johan Montagnat
b
(PhD),
Nicholas Ayache
b
(PhD), Christophe Koehl
a
(E), Olivier Dourthe
b
(MD), Benoit Malassagne
a
(MD),
Michelle Smith
a
(MD), Didier Mutter
a
(MD, PhD), Jacques Marescaux
a
(MD)
Correspondence to Pr. Luc SOLER,
a
IRCAD, 1 place de l’hôpital, 67091, Strasbourg, France
Phone: 33 388 119 065 Fax : 33 388 119 099
Email : Luc.Soler@Ircad.u_strasbg.fr
Key link : www.virtual-surg.com , www.ircad.org
b
Epidaure Project INRIA, BP 93, 06902 Sophia Antipolis France
funding:
• EUREKA Master project of the European Community
• La Ligue contre le cancer, comité du Haut-Rhin
• Région Alsace
• Fondation pour la recherche médicale
• IRCAD, Digestive Cancer Research Institute
Article based on Medical Imaging 2000 Image Processing presentation in San Diego
ABSTRACT
Objective: To improve the planning of hepatic surgery, we have developed a fully automatic anatomical,
pathological and functional segmentation of the liver derived from a spiral CT scan.
Materials and methods: From a 2mm thick enhanced spiral CT scan, a first stage automatically
delineates skin, bones, lungs, kidneys and spleen, by combining the use of thresholding, mathematical
morphology and distance maps. Next, a reference 3D model is immerged in the image and automatically
deformed to liver contours. Then an automatic gaussians fitting on the imaging histogram estimates the
intensities of parenchyma, vessels and lesions. This first result is next improved through an original
topological and geometrical analysis, providing an automatic delineation of lesions and veins. Finally, a
topological and geometrical analysis based on medical knowledge provides hepatic functional
information invisible in medical imaging: portal vein labeling and hepatic anatomical segmentation
according to the Couinaud classification.
Results: Clinical validation performed on more than 30 patients shows that this method’s delineation of
anatomical structures is often more sensitive and more specific than manual delineation by a radiologist.
Conclusion: This study describes the methodology used to create the automatic segmentation of the liver
with delineation of important anatomical, pathological and functional structures from a routine CT scan.
Using the methods proposed in this study, we have confirmed the accuracy and utility of the creation of 3
– dimensional liver model when compared with the conventional reading of the CT scan by a radiologist.
This work, may allow an improvement in preoperative planning of hepatic surgery by more precisely
delineating liver pathology and its relation to normal hepatic structures. In the future this data may be
integrated with computer-assisted surgery and thus represents a first step towards the development of an
augmented reality surgical system.
Keywords: segmentation, gaussians fitting, mathematical morphology, discrete topology, labeling,
hepatic surgery
Keylink: www.virtual-surg.com, www.ircad.org
1. INTRODUCTION
One of the major goals of computerized medical imaging analysis is to automatically detect, identify and
delineate anatomical and pathological structures in 3D medical images. 3D modeling of these structures
then allows for easier and more extensive visualization and exploitation of images. In hepatic surgery,
medical imaging is used to detect and localize hepatic lesions and their relationship to vascular structures,
especially the portal vein that defines the hepatic functional anatomy consisting of several anatomical
segments
1,2
. There are several different definitions for dividing the liver into functionally meaningful
parts that represent the resection unit. Different authors have proposed the division of the liver into two
hemilivers, or into four segments based on the Goldsmith and Woodburne definition
3
or into eight sub-
segments based on the Couinaud definition
4
which is today considered the international standard
1
.
In order to detect lesions and to localize vascular networks defining the anatomical segments,
radiologists currently use helical Computed Tomography scan images with intravenous contrast infusion
(helical CTI). In these images, tumors appear as dark nodules within bright hepatic tissues whereas vessel
trees appear as a network brighter than the liver parenchyma. However, detection of the lesion or
localization of the vessels is often difficult to process due to a variable image contrast between liver
parenchyma and vessels, and also due to an important image anisotropy, the slice thickness being three
times larger than the pixel width.
Therefore in hepatic surgery, one of the goals of computerized medical imaging processing is to
automatically delineate liver, lesions, vessels and anatomical segments from the imaging studies. Several
authors proposed to delineate the liver contours from CTI images with an automatic
5,6,7,8,9
, or semi-
automatic process
10
. Several methods use a deformable model, either to directly delineate structures
5,7
, or
to improve the results of a previous delineation technique
6
. In addition, vascular tree segmentation has
been performed in different studies
11,12,13,10
. Among these works, the method of Zahlten et al.
12,13
allows
extraction of the portal vein from abdominal CT-scan images, using a region growing technique. This
technique has the advantage of giving a topological information about the venous tree, which is useful for
building all anatomical segments
14
. However, since it requires a manually-set threshold and an initial seed
point, this technique is not fully automatic. Finally, there have been very few studies
15,16
about the hepatic
lesion delineation, sometimes performed by the same methods used to isolate other anatomical
structures
7
.
Among all these studies, the work of Gao et al.
6
is best suited for hepatic surgery planning since it
provides a general solution allowing the delineation of the hepatic anatomy, even if the vascular system
may not be clearly delineated. But this method of liver segmentation does not provide good results with a
liver which contains a large sub-capsular tumor. Also, the work of the MEVIS team
12,13,14
performs portal
vein labeling and anatomical segments delineation, but it always reconstructs eight sub-segments even if
the patient has a different number of segments. Moreover, this segmentantion technique requires many
time-consuming interactions.
In this article, we propose an original three step anatomical segmentation method, based on the
translation of anatomical knowledge into topological, geometrical and morphological constraints. This
method allows thus for automatic extraction of liver, hepatic vessels, hepatic lesions and also of the
anatomical segments with respect to the three most common definitions: hemilivers, Goldsmith and
Woodburne definition and Couinaud definition.
2. AUTOMATIC LIVER, LESIONS AND VEINOUS SYSTEMS DELINEATION
2.1. Patients dataset
This study has been performed on a set of 35 CT-scans with slices from 2 mm to 3 mm of thickness,
acquired after contrast agent injection at portal phase, from an helical Siemens Somatom 4 plus CT-scan.
The database is composed of 33 images with intravenous injection, and two portoscans. It includes
healthy subjects, patients with lesions (cyst or tumors), and patients after segmentectomy. Furthermore,
the rate of contrast product infiltration into hepatic venous systems is quite variable from one patient to
another, due to a difficult evaluation of the portal time.
2.2. First stage: skin, lungs, bones, kidneys, spleen and liver delineation and image improvement
This first stage of our method automatically extracts step by step, the skin, lungs, bones, kidneys, the
spleen and the liver of a patient, from a CT-scan image. Our method consists in translating anatomical
information obtained by the medical imaging and transforming this information by the way of several
simple intensities, morphological, topological and geometrical constraints. The intensity in Hounsfield
units of air, fat tissue, water and bones are known and are respectively -1000 HU, -120 HU to -80 HU, 0
HU, and 500 HU to 3000 HU. Air is mainly outside the patient and in the lungs (some air may be
eventually found into the digestive system too). Isolating the air allows us to easily extract the skin and
the lungs boundaries. A simple threshold does not allow for isolating the bones. Because of the contrast
agent, others structures, such as the aorta, appear bright. To overcome this, we first isolate the fat tissue
(thresholding followed by morphological operation). The bones are then characterized as the brightest
structures close to the fat tissue.
Kidney and spleen delineation is more difficult due to their intensity variation. We then propose a
solution based on the gray-level histogram analysis of the image limited to regions including the spleen
and kidneys. Indeed, the right inferior quarter of the image contains essentially a part of the liver and the
right kidney, whereas the left inferior quarter of the image contains only the right kidneys and the spleen.
Thus, a comparative analysis of the gray-level histograms allows us to find the intensity range of kidneys,
spleen and liver parenchyma, identically localized on both histograms. We then delineate the kidneys and
the spleen by performing a thresholding followed by morphological operators.
After all of these anatomical structures are removed from the original image, we finally extract liver.
From several existing methods, we chose to use the Montagnat and Delingette method
5
who proposed an
hybrid deformation framework that consider the global transformations computed in the registration
framework
17
as a deformation field similar to the local deformation field of the deformable models
18,19
scheme. This method applies thus to each S(i) vertex of the model with a locality parameter l, a combined
force f(i):
f(i) = (1-l) * GlobalForce(i) + l * LocalForce(i) (1)
It is possible to apply to the model in this single framework completely local (l=1), completely global
(l=0) as well as any intermediate (0<l<1) force between those two ends. This framework introduces a
global constraint in the deformation process that may be scale through the l parameter. It makes models
more reliable: they are less noise and outliers sensitive. Moreover, the geometric quality of meshes
produced by the deformation scheme is better.
Application of this method to liver delineation first requires the initialization of a 3D reference model
in the image. In order to obtain a liver contour by deformation, it is easier and more reliable to use an
initial model with a similar shape. This model is a liver template computed onto visible human data of the
National Library of Medicine. Delineation is then composed of several stages. Fig. 1 and 2 respectively
represent the liver template being bent along each stage and its cut superimposition onto on of the CT
scan slides.
(a) initialization (b) rigid and affine registration
(c) hybrid deformation l low (d) hybrid deformation l high
Figure 1. Shape evolution of the model along stages.
(a) initialization (b) rigid and affine registration
(c) hybrid deformation l low (d) hybrid deformation l high
Figure 2. Evolution of the model cut on one slice of the 3D image.
From the resulting liver delineation, we chose to reduce and improve the initial image in order to speed
the process and also to improve the lesions and vessels delineation. Firstly, the extracted liver is used as a
mask, which reduces the initial image to the region of interest of the liver. Secondly, the reduced image is
filtered with the anisotropic diffusion detailed in
20
. It then reduces the textured aspect of CT scan without
loss of structure borders. As shown on Fig. 3, the textured aspect of the initial image is changed in
homogenous intensity areas, whereas borders separating parenchyma, vessels and dark areas are
preserved.
Figure 3. Reduced image before and after anisotropic diffusion, with zoom on two areas: area 1 (up) and area 2 (down)
2.3. Second stage : automatic delineation of lesions and vascular systems
We saw previously that Gao et al.
6
proposed a classification method of all internal structures of the liver.
To do this, the authors estimated the intensity distributions of three tissue classes, lesions, parenchyma,
and vessels, as trapezoidal functions and used the percentage of voxels belonging to each class for the
visualization. We chose to modify their method by considering that the distributions of the same three
tissue classes follow a normal law, these distributions being then used to consider thresholds allowing
segmentation for each structure. The fitting of distribution onto the gray level histogram is performed by
the Levenberg and Marquardt’s method
21
, which minimizes a least square criterion and which is currently
used for other organs in many articles
22,23,24,25
.
In the liver case, the major limitation of this method is the need to obtain a good initialization of the
distribution parameters whereas only the peak corresponding to the liver is usually visible. To fill this
handicap, we propose an original resolution in two stages. The first stage fits, on the gray-level histogram,
the gaussian curve that corresponds to the parenchyma whose peak is always visible. The subtraction of
the resulting gaussian from the histogram provides thus distribution of points that do not belong to the
liver (with the some errors close to the first adjustment). From this subtraction, the second stage initializes
the two last gaussians and fits the three class gaussians on the initial gray-level histogram. The thresholds
are then estimated as the intensities for which two neighbor gaussians cross, defining thus S
LF
, the
threshold separating the lesions voxels from the liver voxels, and S
FV
, the threshold separating the liver
voxels from the vessels voxels (Fig. 4).
