fully automatic liver segmentation for contrast-enhanced...

8
Fully automatic liver segmentation for contrast-enhanced CT images aszl´ o Rusk´ o, Gy¨ orgy Bekes, G´ abor N´ emeth, and M´ arta Fidrich GE Hungary ZRT. Healthcare Division, Akron u. 2., H-2040 Buda¨ ors, Hungary [email protected] Abstract. The need for fast and precise segmentation has increased recently due to the spread of systems for computer aided diagnosis and therapy planning. The manual segmentation of the liver is very time con- suming, so it is desired to develop a method that can precisely segment the liver without any human interaction. In this paper we propose a fully automatic method for liver segmentation on contrast-enhanced (por- tal venous) CT images. Our method is essentially an advanced region- growing the result of that is improved by various pre- and post-processing steps, like intensity based ROI detection, separation of liver and heart, additional segmentation at the right lung lobe, vessel (IVC) removal, and cavity filling. According to our experiments the method can efficiently segment the liver parenchyma in many cases, however, in some cases the result may exclude very large lesions. 1 Introduction Computer assisted planning of various liver treatments (minimally invasive ther- apies, oncology liver sectioning, living donor transplantation) is primarily based on computed tomography (CT), which can be an important aid for operability decisions and visualization of individual patient anatomy in 3D. The planning is based on the liver volume, the anatomical liver segments, the vessel structure, and the relation of lesions to these structures. The detection of the boundaries between the segments is then the first step of the preoperative planning. Ra- diologists currently use CT images with intravenous contrast infusion, in order to detect lesions and vessels in the liver. The key point of the above-mentioned treatments is the liver volume segmentation. This step is quite time consuming when it is done manually. Our aim is to develop a method that is precise, quick and robust enough to use it in the every day clinical practice. There are several published methods about segmentation of CT images. Most of these methods are some variants of the region-growing, active contour/surface, level-set, or thresholding, classification algorithms (of course, all of them are adapted to the specific situations and they are equipped with several pre- and post-processing operations). In addition, the methods are often based on some statistical, anatomical, or geometric model. Soler [1] proposes a fully automatic method to segment the liver from contrast- enhanced CT scans. This method delineates the skin, bones, lungs, kidneys and T. Heimann, M. Styner, B. van Ginneken (Eds.): 3D Segmentation in The Clinic: A Grand Challenge, pp. 143-150, 2007.

Upload: letuyen

Post on 28-Jul-2018

220 views

Category:

Documents


0 download

TRANSCRIPT

Fully automatic liver segmentationfor contrast-enhanced CT images

Laszlo Rusko, Gyorgy Bekes, Gabor Nemeth, and Marta Fidrich

GE Hungary ZRT. Healthcare Division, Akron u. 2., H-2040 Budaors, [email protected]

Abstract. The need for fast and precise segmentation has increasedrecently due to the spread of systems for computer aided diagnosis andtherapy planning. The manual segmentation of the liver is very time con-suming, so it is desired to develop a method that can precisely segmentthe liver without any human interaction. In this paper we propose a fullyautomatic method for liver segmentation on contrast-enhanced (por-tal venous) CT images. Our method is essentially an advanced region-growing the result of that is improved by various pre- and post-processingsteps, like intensity based ROI detection, separation of liver and heart,additional segmentation at the right lung lobe, vessel (IVC) removal, andcavity filling. According to our experiments the method can efficientlysegment the liver parenchyma in many cases, however, in some cases theresult may exclude very large lesions.

1 Introduction

Computer assisted planning of various liver treatments (minimally invasive ther-apies, oncology liver sectioning, living donor transplantation) is primarily basedon computed tomography (CT), which can be an important aid for operabilitydecisions and visualization of individual patient anatomy in 3D. The planning isbased on the liver volume, the anatomical liver segments, the vessel structure,and the relation of lesions to these structures. The detection of the boundariesbetween the segments is then the first step of the preoperative planning. Ra-diologists currently use CT images with intravenous contrast infusion, in orderto detect lesions and vessels in the liver. The key point of the above-mentionedtreatments is the liver volume segmentation. This step is quite time consumingwhen it is done manually. Our aim is to develop a method that is precise, quickand robust enough to use it in the every day clinical practice.

There are several published methods about segmentation of CT images. Mostof these methods are some variants of the region-growing, active contour/surface,level-set, or thresholding, classification algorithms (of course, all of them areadapted to the specific situations and they are equipped with several pre- andpost-processing operations). In addition, the methods are often based on somestatistical, anatomical, or geometric model.

Soler [1] proposes a fully automatic method to segment the liver from contrast-enhanced CT scans. This method delineates the skin, bones, lungs, kidneys and

T. Heimann, M. Styner, B. van Ginneken (Eds.): 3D Segmentation in The Clinic:A Grand Challenge, pp. 143-150, 2007.

spleen, by combining the use of thresholding, mathematical morphology and dis-tance maps then the liver is extracted. A 3D reference model is generated frommanually segmented livers and adjusted onto the image with rigid and affineregistration. The model is deformed to get the final result. The weakness ofthis method is that one phase is used, which is acquired according to a specialprotocol and does not correspond to the general practice.

An automatic approach for segmentation of the liver from CT images basedon a 3D statistical shape model is presented in the paper of Lamecker [2]. Thisiterative technique first builds a statistical model from a training set of shapes.Each shape is defined by some anatomically specific points sampled on the sur-face. The next step is the positioning the mean shape into the image. Then singleshape adjustment is applied. Unfortunately, there is no clinical evaluation andthe selection of the landmarks is very difficult due to the very variable shape ofthe liver.

The level-set method-family ([3], [4], [5]) has been successfully used for med-ical image segmentation. The advantages of the level set approach are that ithandles topological changes and defines the problem in one higher dimension.The main disadvantages are that these methods are time-consuming and theyusually produce over-segmentation.

The active contour [6] method is used to segment abdominal organs in theclinical practice. It works well on native images, because the organs are homo-geneous. In case of contrast-enhanced images, the contrast agent is cumulateddifferently in different parts of the liver. For example the vessels and some tu-mors will have higher intensity than the liver parenchyma. The active contourmethods starts from a smaller region and try to blow it up and fit the surface tothe contour of the organ. The vessels and tumors set back the regular growingof the surface.

The region-growing based approaches [7] can provide good results on contrast-enhanced images. Such method starts from a small region (environment of inputcurve, or point), and each neighboring voxel is added to the actual region, ifits intensity is corresponds to a pre-defined range. The region-growing can effi-ciently close round the vessels and tumors (in contrast with active contour), butit is very sensitive to its input and can easily flow into neighboring organs thathave similar intensity.

In the next chapter we describe a fully automatic segmentation methodfor liver segmentation. Our method is based on intensity analysis and region-growing, where the issue of under- or over-segmentation is handled in variousways. In Section 3, we present the evaluation of this method using the test examset provided by the MICCAI Grand Challenge on clinical liver segmentation.

2 The method

Our method consists of the following main steps. First a seed region is determinedthat involves voxels which are located inside the liver. Then, the liver is separatedfrom the heart to prevent over-segmentation in this region. Starting from the seed

144

region the liver is segmented using an advanced region-growing (RG) method.The segmentation is followed by various post-processing steps, so that the sizeof under- or over-segmented regions is reduced significantly.

2.1 Determine seed region

When the CT image to be segmented is enhanced using some contrast agent,the abdominal organs can be easier separated due to different contrast intakeof the different organs, which can be exploited when the region of the liver isautomatically determined. Besides the characteristic contrast intake, the size(largest organ) and the location (mostly on the right side) of the liver make iteasier to determine significant portion of its volume without user interaction.

In this paper we assume that image to be segmented is a contrast-enhancedCT scan of the liver that is acquired in the portal venous phase (contrast agent isvisible in the portal vein). According to several exams, which were acquired usingdifferent types of scanners and contrast agents, with different image resolutionand quality, the intensity of the liver voxels is in the range of [-50,250] HU. Inorder to determine the mean intensity of the liver voxels for a particular exam,a histogram can be calculated. If reliable result is needed, only those voxels areincorporated, which are located in the right side of the body and the intensity ofwhich is in the range [-50,250] HU (air, fat, bones are excluded). After smoothingthis histogram, it has two significant maxima. One of them belongs to the musclesthe other belongs to the liver. Since the intensity of the liver is always higherthan that of the muscles, the maximum greater than 80 HU represents the modusof liver intensities in all cases.

After this modus is determined, the rough estimation of the minimal andmaximal intensity of liver voxels can be easily calculated based on the histogram,which makes it is possible to find a region inside the liver automatically. First,a binary image is created based on the original grayscale image such that thevoxels, the intensity of which is in the previously determined range, have value1, and all other voxels are zero valued. This image usually consists of severalregions, the intensity of which is similar to the liver’s intensity.

The image of possible liver regions is then eroded so that small regions aredeleted. Since the liver has the largest compact volume in the abdomen, a spherewith a relatively large radius can be used for erosion. The value of this radiuswas determined based on several exams. After eroding the image, the largestconnected region is considered as seed region for the segmentation. Although,the size of this region may vary among the different exams, this method providesa reliable set of liver voxels in all cases. The average size of this region for 20training exams was 15% of the total liver volume.

2.2 Liver heart separation

In the image belonging to the portal venous phase the liver and the heart hasnearly the same intensity, so the result of a 3D RG method usually involves theheart. Since large over-segmented regions shall be eliminated, it is important

145

to keep the RG off the heart. The liver-heart separation algorithm takes theadvantage of the anatomical feature that the bottom of the lung fits the liversurface. We separate the heart from the liver by means of connecting the bottomof the left and right lung lobes with a surface. The method first determines thebottom of the right and left lung lobes. Then, for each coronal slice a minimal-length curve is found, which connects the two lungs and goes along large gradientvalues. The set of these curves defines a surface that is used to prevent the RGto go into the heart.

In case of abdominal CT scans, slices at the top of the image include thebottom of the lung (if not, there is no need for liver heart separation anyway).Starting from the topmost slice both lung lobes can be segmented based on thecharacteristic intensity of the air. In order to find seed-points for the left and theright lung lobes in the topmost slice, we do the following. First we determine thelargest connected region, where the intensity of voxels is higher than -400 HU.This is the body region. Inside this region we determine the largest connectedair region for the left and the right side, separately. Using these regions as seed,a 3D RG method can segment the left and right lung lobes.

After the lung lobes are segmented, the coronal slices of the CT image areprocessed. The goal is to determine two curves representing the bottom contourof the right and the left lobes and connect the leftmost point (L) of the rightcurve and the rightmost point (R) of the left curve. When L and R are availablefor each coronal slice, we use the following method to connect them. Startingfrom L we try to reach R such that we encounter voxels with the largest possiblegradient value. Going from right to the left, in each step we choose the locationof the largest gradient value found in the local environment of the previous point.When we reach the vicinity of R we connect the current point with R with adiscrete line.

After this curve is determined for each coronal slice the surface separating theliver and heart is calculated by averaging the curves located in the neighboringslices, which provides a smoother solution. Finally, for each coronal slice thevoxels, which are located above the surface, are set to an artificial intensityvalue (3000) so that the RG cannot go into this region.

2.3 Region-growing

The liver parenchyma is nearly homogeneous, so a RG method can efficientlydetermine most of the liver volume. In order to use a RG method we need someseed-points and the intensity range of the voxels belonging to the liver. The set ofinitial points is determined according to Section 2.1, while the intensity intervalis calculated in the following way. First, we create the histogram of intensitieslocated in the environment (of 5 mm radius) of all initial points. Then, wecalculate the intensity interval based on the modus of this histogram, and theleft and right standard deviation of intensities with respect to the modus.

According to our experiments the intensity range of the liver voxels cannot beperfectly determined. If the RG uses a global intensity range some regions can beunder- or over-segmented. Thus, it is important to find different ways to correct

146

the segmentation result in these regions. In order to reduce the probability ofover-segmentation we made the following modification to the RG method. Duringa 3D RG method a voxel is added to the region if all voxels in its neighborhoodhave acceptable intensity. In the literature usually 6-neighborhood is used. Ourexperiments showed that using larger neighborhood (e.g. sphere with 5 mmradius) reduces the probability of over-segmentation significantly. When sucha large radius is used for RG, the neighborhood consists of a large number ofvoxels. Due to the noise, we have to use some tolerance, when we check if thevoxels’s environment satisfies the intensity condition or not.

Even though the noise is reduced (using any filter), under-segmentation mayconcern several region of the liver, which are less homogeneous. According to ourexperiments, the region near the lung, some lesions, and vessels, the intensityof which is significantly lower or higher than the intensity of the normal liverparenchyma may be under-segmented. In the following subsection we discusshow these problems can be eliminated.

2.4 Post-processing

In some cases the liver is under-segmented near the right lung lobe, where theratio of intensities lower than the minimal intensity exceeds the tolerance. Thisproblem can be corrected by additional segmentation that allows lower intensityrange in the region located between the surface of the segmented liver and rightlung lobe. First, we determine the surface voxels for the right lung and calculatethe surface normal vector for each of them. If the normal vector of a surfacevoxel points toward a liver voxel that is closer than a predefined distance, thesurface voxel is marked. In the next step we connect each marked lung surfacepoint with the corresponding liver voxel using a discrete 3D line. Then, thelocal environment is calculated for each line, which defines a closed connectedregion between the liver and the right lung lobe. We calculate a new intensityinterval based on this region, which is used by an additional RG. This methodstarts from liver surface points and limited only to this region, so it cannot causeover-segmentation in other part of the liver.

In case of portal venous images the intensity of the inferior vena cava (IVC)is very similar to the intensity of the liver parenchyma. Since the diameter of thisvessel is significant (20-30 mm), the neighborhood connected RG method leaksout through the IVC in nearly 40% of the cases. In order to remove IVC fromthe segmentation result we do the following post-processing. Our main idea isto detect those parts of the segmented liver, which are similar to a cylinder witha specified diameter. Since this vessel is vertically oriented, the cross section ofthe IVC is a circular region in each axial slice. In order to detect circles with agiven diameter the circular Hough transform is used. In our case, however theradius of the circle varies, so instead of a circle we use a ring such that theinner radius is smaller and the outer radius is greater than the average radiusof IVC. Using this ring the entire 3D contour of the liver is processed, so thata probability map is created, where higher values represent voxels, which arelocated inside a horizontal tube with radius similar to IVC. After thresholding

147

the probability map, we process it slice by slice. We locate all local maxima inthe slice, and for each maximum we check whether a closed contour is foundaround the location of the maximum inside its local environment. If so, we markthe region around the given maximum in the segmented image as candidatefor erasing. In the next step we determine the largest connected region in thesegmented image that consists of unmarked liver voxels. All liver regions, whichare not connected to the largest one, are also marked as candidate for deleting(such region can be found along the IVC, where the vein has a branching point).Finally, all candidate region is erased, the vertical length of which is greater thana predefined constant (so that we don’t erase the bottom peak of the right andleft liver lobes).

An intensity-based segmentation will not involve voxels belonging to the vas-cular structure (portal vein), which have higher intensity. In the clinical practice,however, the vessel is considered as part of the liver as long as it is completelysurrounded by liver parenchyma. In order to fill the vessels inside the liver wedo the following. We determine the contour of the segmented liver, and calculatethe surface normal for each of them. Then, we mark a surface voxel if there isanother liver voxel in the direction of surface normal, such that its distance fromthe corresponding surface point is nearly equal to the average diameter of portalvenous branches. Finally, we dilate the liver at each marked surface voxel usinga sphere the radius of that is equal the average radius of the portal vein. Herewe note, that smaller tumors, which form closed cavity inside the liver volume,can be filled using any simple cavity filling method.

According to our experiments, the RG does not need the image to have highresolution. In order to speed up the segmentation we omit slices such that theslice thickness is between 2 and 3 mm. After the segmentation, the result isneeded in the resolution of the input image. In order to get a smooth interpo-lation between the segmented slices, we create the triangular representation ofthe liver surface, we smooth it, and convert it to voxel image with the resolu-tion of the input image. We note, that this surface representation allows furtherrefinement of the results based on the image gradient, which have not beenimplemented in this work yet.

3 Results

The organizers of MICCAI Workshop on 3D Segmentation in the Clinic haveevaluated our method on 10 test exams, which involves easy, average, and difficultcases. Figure 1 shows our result for each of these cases. Based on the images, wecan claim that our method performs well, if the liver does not have very largelesions. In other cases (depicted in the middle and the bottom rows), however, theliver can be very under-segmented, which is the most important problem we haveto solve in the future. Here we note, that our method was primarily developed toaid minimally invasive liver therapies (e.g. embolization of tumors), where suchlarge lesions are not considered.

148

Fig. 1. From left to right, a sagittal, coronal and transversal slice from a relatively easycase (1, top), an average case (4, middle), and a relatively difficult case (3, bottom). Theoutline of the reference standard segmentation is in red, the outline of the segmentationof the method described in this paper is in blue. Slices are displayed with a window of400 and a level of 70.

Table 1 displays the numerical evaluation of our results with respect to 5important metrics. When the liver has large lesions (3, 4, 8, 10) our methodprovides bad result concerning all metrics because the large lesions are under-segmented. Although, our method is based on intensity analysis, it is not sensitiveto the majority of lesions, the size of which allows of embolization or ablation.These lesions, which have significantly higher or lower intensity than the liver,are filled by the post-processing steps in most of the cases. Focusing on thesecases, (1, 2, 5, 6, 7, 9) our method performs an average 76 of total score, whichis a bit better than a non-expert manual segmentation. Our method providesthe result in 56 seconds in average (min 31, max 113) for the 20 training examsusing an Intel Pentium 4 CPU 3GHz processor. Considering that an average39% of this time is spent with IVC removal (that was not optimized yet) thismethod can efficiently segment the liver volume in most of the clinical cases, butit definitely needs further improvements to handle the extreme cases.

149

Dataset Overlap Error Volume Diff. Avg. Dist. RMS Dist. Max. Dist. Total[%] Score [%] Score [mm] Score [mm] Score [mm] Score Score

1 7.0 73 1.1 94 1.1 73 2.0 73 18.1 76 782 7.1 72 4.5 76 1.1 73 2.2 70 22.7 70 723 20.6 20 -16.0 15 3.9 2 9.0 0 50.0 34 144 14.1 45 -5.9 69 2.6 35 6.2 14 45.1 41 415 8.4 67 -2.4 87 1.5 63 2.8 61 28.6 62 686 7.2 72 1.2 94 1.1 73 2.0 73 20.2 73 777 5.8 77 0.7 96 0.9 78 1.7 77 17.4 77 818 10.8 58 -9.3 50 1.9 52 3.8 47 26.0 66 559 6.9 73 -0.5 98 0.9 78 1.7 77 15.0 80 81

10 19.7 23 -16.6 12 3.3 17 7.0 3 40.1 47 20Average 10.7 58 -4.3 69 1.8 54 3.8 50 28.3 63 59

Table 1. Results of the comparison metrics and scores for all ten test cases.

Acknowledgment

Some parts of the presented method were developed in cooperation of GE Hungary and

the University of Szeged. Hereby, we would like to thank the university team members,

namely Norbert Bara, Csaba Domokos, Krisztina Dora, Tamas Korodi, Laszlo Reszegi,

Arpad Tigyi, and Norbert Zsoter for their contribution. We are also very grateful to the

medical evaluation team lead by Prof. Andras Palko Ph.D, namely Katalin Gion M.D.,

Edit Kukla M.D, and Endre Szabo M.D. for their very important clinical feedbacks.

References

1. Soler, L., Delingette, H., Malandain, G., Motagnat, J., Ayache, N., Koehl, C., Dour-theb, O., Malassagne, B., Smith, M., Mutter, D., Marescaux, J.: Fully automaticanatomical, pathological, and functional segmentation from ct. Proc. SPIE MedicalImaging 3979 (2000) 246–255

2. Lamecker, H., Lange, T., Seebass, M.: Segmentation of the liver using a 3d sta-tistical shape model. Technical Report ZIB-Report 04-09, Konrad-Zuse-Zentrum frInformationstechnik Berlin (April 2004)

3. Sethian, J.A.: Level Set Methods and Fast Marching Methods. Cambridge Univer-sity Press (1999)

4. Shi, Y., Karl, W.C.: A fast implementation of the level set method without solvingpartial differential equations. Technical report, Boston University, Department ofElectrical and Computer Engineering (January, 2005)

5. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. International Journalof Computer Vision 22(1) (1997) 6179

6. Bekes, G., Nyul, L.G., Mate, E., Kuba, A., Fidrich, M.: 3d segmentation of liver,kidneys and spleen from ct images. Proc. International Journal of Computer As-sisted Radiology and Surgery 2(1) (2007) 45–46

7. Pohle, R., Toennies, K.D.: Segmentation of medical images using adaptive regiongrowing. Proc. SPIE Medical Imaging 4322 (2001) 1337–1346

150