Automatic Segmentation of Liver From CT Images Using Probabilistic Atlas and
Template Matching
Yen-Wei CHENa, b 1, Amir H. FORUZANb, c, Chunhua DONGb, Tomoko TATEYAMAb, Xianhua HANb
aCollege of Computer Science and Information Technology, Central South Univ. of Forestry and Technology, Hunan, China
b College of Information Science and Engineering, Ritsumeikan University, Shiga, Japan
cDepartment of Biomedical Engineering, Engineering Faculty, Shahed University, Tehran, Iran
Abstract. A framework is proposed for automatic liver segmentation from CT volumes using probabilistic atlases and template matching techniques.
Probabilistic atlases of human anatomy have been widely used for organ segmentation, which is used as a prior probability in a Bayes framework. The challenge is how to register the atlas to the patient volume. In this paper, we propose a template matching based technique for probabilistic atlas based organ segmentation. In our proposed method, we first find a Region of Interest (ROI) of the organ, which is based on human anatomic structure, and then the probabilistic atlas is used as a template to find the organ in the ROI by the use of template matching.
Keywords. Liver segmentation, CT volume, Probabilistic atlas, Template matching, Region of interest (ROI), Bayes theory
Introduction
Evaluation of liver geometry, its vessels structures, liver’s tumors sizes and locations are considered as a critical step prior to liver treatment planning [1, 2]. The initial stage of any CAD/CAS system that deals with liver is segmentation. A wide range of image processing techniques have been used by researchers to develop liver segmentation algorithms. Most of proposed methods, such as region growing [3], clustering [4], active contours [5-7], graph cut [8-10], are based on intensity information and do not use any anatomical information. Their performance is not good enough when the image contains noise and the contrast between objects and background is low.
Recently, a new trend is to use anatomical models, such as probabilistic atlas [11, 12], statistical shape model [13, 14] for organ segmentation. In anatomical model based methods, the anatomical model can be used as a prior location and shape information of organs. Thus, these methods are robust against noise, however, they are sensitive to initialization of their parameters such as pose and initial contour. The challenge is how to register the atlas to the patient volume. This paper is focused on probabilistic atlas based liver segmentation, which is usually used as an initial segmentation process. In
1 Corresponding Author: Yen-Wei Chen ([email protected])
R. Neves-Silva et al. (Eds.) IOS Press, 2014
© 2014 The authors and IOS Press. All rights reserved.
doi:10.3233/978-1-61499-405-3-412
conventional probabilistic atlas based segmentations [11, 12], the diaphragm or the spine is used as a landmark, which can be automatically extracted based on their intensity information. With these landmarks, the transformation between the reference volume and the patient volume can be estimated by the use of thin plate spline or affine registration techniques. Then the probabilistic atlas can be transformed (registered) to the patient volume. There are two limitations of the conventional method: (1) Errors of landmark extraction may cause incorrect estimation of the transformation; (2) Even though we can estimate the transformation of landmarks accurately, it is difficult to estimate the accurate transformation of the liver because of the large variation of anatomical structure. In this paper, we propose a template matching framework for probabilistic atlas based organ segmentation. In our proposed method, we first find a Region of Interest (ROI) of the organ, which is based on human anatomic structure, and then the probabilistic atlas is used as a template to find the organ in the ROI by the use of template matching,
The paper is organized as follows: In section 1, we describe our proposed method including probabilistic atlas construction, bounding box setting, template matching.
Section 2 shows experimental results and section 3 concludes the paper.
1.Materials and Methods
The flowchart of our proposed method is shown in Figure 1.
Figure 1. The flowchart of the proposed method.
The proposed method segments an abdominal organ based on a set of prior information. Prior information includes the approximate location of an organ, histogram of the organ’s intensities, and a probabilistic atlas of the organ. The approximate location of an organ is employed to estimate the ROI of the organ. The
histogram of the organ’s intensity is used to build the likelihood image. Then, the likelihood image together with the probabilistic atlas are employed in template matching technique. The template matching is proposed to estimate an initial segmentation of the organ. To find the final segmentation result, the level-set algorithm is used.
In the following sections, we describe how to segment liver in the abdominal region. However, the algorithm can be extended to segment other organs.
1.1. Bone Extraction and Organ ROI
The basic idea of our proposed method is to find an organ ROI based on anatomical structure firstly and then find the organ in the ROI by using template matching. We use bone as a standard reference to estimate the ROI. Several methods have been proposed for bone extraction [15, 16]. The main approach of these methods has been to estimate the intensity range of the bone and apply the region-growing technique to find the bones. We follow the same approach and find the intensity range of bones, employ region-growing algorithm to segment skeleton, and use morphological operators to refine the results. The flowchart of our bone extraction is shown in Fig.2. One typical extracted bone is shown in Figure 3(a).
Figure 2 The flowchart of our bone extraction method.
Because of the variations in anatomical structure, the positions of organs will differ for every sample. We use training samples to build the organ ROI (bounding box). We randomly select one sample as a reference volume and other samples are normalized to the reference volume by affine transformation with extracted bones. One example of bone registration is shown in Figure 3(b). The gray is the reference bone, the blue is the moving bone before registration and the red is the moving bone after registration.
We use the transform matrixes, which are estimated in the registration of a moving bone to reference bone, to register the organs of other sample to the reference organs.
After the organs of all training data are registered, they build an ensemble, together with the organ of the reference data. The ROI of the ensemble of livers is shown in Figure 3(c).
Figure 3 (a) Examples of bone extraction, (b) registration of bone, and (c) the ensemble of eight registered left kidneys with its ROI.
1.2. Construction of Probabilistic Atlas
The organ probabilistic atlas can serve as a prior probability map of the organ. We used the segmented livers in the training data to build a probability atlas for the liver.
Registration of the liver masks is the first step in atlas building, which is performed using a rigid scheme. In this step, we do not use any bone as the reference image and we use the same data mask as bone reference data as liver reference mask. The remaining masks are considered as the moving images and are registered to the reference mask. The ROI is found in the world coordinate since it should be used in the registration algorithm. When input masks are registered, we select the largest region containing all the input masks. An important point in finding the ROI of the masks is to consider the direction cosines of the input data. The matrix of direction cosine is stored in the header file of a MHD image as the “TransformMatrix” field. To accurately build the atlas, we cannot add the two mask data pixel-wise since the spacing and origin of input data are not the same. After finding the largest region, which contains all livers in the world coordinates, we resample all masks using the minimum input spacing and put the origin of all resampled data to (0, 0, 0). Now, all masks occupy the same physical location in the world coordinates and have the similar spacing and origin. Now, pixel- wise addition of resampled input masks makes meaningful. The result of summation is an image, which is later scaled by 1/number_of_input_masks. Thus, the probability atlas is built, which is shown in Figure 4.
Figure 4 Probabilistic atlas of the liver.
1.3. Intensity Based Probabilistic Map
Intensity based probabilistic map (or likelihood map) of an organ refers to an image in which the intensity of a pixel corresponds to the probability of belonging to that organ. Thus, the input to this stage is the original image and the output is a probability image. For example in case of liver, the intensities of the pixels in the original image are converted into a probability using the histogram of the liver. Typical histogram of the liver is shown in Figure 5(a), which can be considered as likelihood of the liver.
The original CT image and the corresponding intensity based probabilistic map are shown in Figure 5(b) and 5(c), respectively. Comparison of Figure 5(b) and 5(c) reveals that the liver can be more easily distinguished from other tissues in the intensity based probabilistic map image. But it can also be seen that distinguishing liver from several tissues with similar intensity is a little difficult if we only use intensity information (likelihood).
1.4.Template Matching
Template matching is one of the key innovations of our algorithm. The goal of this step is to build a probability map, which assigns to each pixel the probability of belonging to the organ of interest. To build such a map, we use two probability images, an intensity based probabilistic map (or likelihood map) and a probability atlas. We use prior information of the organ of interest to restrict the processing region in template matching. The template matching algorithm moves the probability atlas in the ROI of the organ and multiplies the probability atlas by the likelihood map. The summation of probabilities of pixels of the result is used as an index and the position of the maximum index is used as the best pose of the initial estimate of the organ. The output of the template matching which corresponds to the maximum index is used to find the initial
segmentation of the organ using a thresholding technique. The final result is obtained through post processing and refinement by the level-set algorithm.
The template matching is based on the idea that those pixels, which belong to liver have similar intensities to the liver’s histogram and their locations are in parts of the image where liver may be found with a higher probability, which is defined as Eq. (1):
) ( )
| (
)
|
(Liver Intensity P Intensity Liver P Liver
P ∝ ×
(1)
In Eq.(1), P(Intensity | Liver) is the likelihood function (likelihood image) which can be obtained using the liver histogram and P(Liver) is estimated by the liver probability atlas. The likelihood image is the same as intensity based probabilistic map described in previous step. However, the size of likelihood image and the probability atlas are not the same. The probability atlas is usually smaller than the intensity based probabilistic map. Starting from an ROI in which the liver is located, we put the probability atlas at the beginning of the region and multiply by intensity map. The result is a probability image. The summation of the probabilities gives us an index of the total probability. This index is maximized where both the likelihood and the probability atlas are in accordance with each other (Eq.(2)).
(i
max,jmax,kmax)=
i,j,k
arg max
P(Intensity|Liver)×Ti,j,k(P(pixels
∑
Liver))
(2)
In Eq.(2), T(.) is the translation of the liver probability atlas. The indices (i, j, k)max
give the best pose for the probability atlas. We translate the probability atlas by (i, j, k)max and multiply the two images to get an estimate of the probability of pixels belonging to liver. One example is shown in Figure 5(e). In order to make a comparison, we also show the result using conventional method, in which the transformation between the reference bone and the patient bone is directly applied to the atlas, in Fig.5(d). As well as Figure 5(c), the white regions have a higher probability of belonging to liver. If we threshold the output of template matching, it will give us an initial segmentation of the liver. Compared with Figure 5(c), we can see that the false positives (other tissues with similar intensity) can be significantly reduced by the use of atlas. It can also be seen that there are still a lot of false positives in the conventional method even we use a probability atlas, because we can not align the atlas to the patient volume accurately, while our proposed method can archive more accurate alignment.
1.5. Final Liver Segmentation
The final step of organ segmentation includes thresholding the output of template matching, smoothing the results by opening morphology operator, use it as an initial organ segmentation, and tuning the initial result by a level-set algorithm [7].
2.Results
We apply our proposed method to real CT volumes. The segmentation results of two typical CT volumes are shown in Figure 6. The red ones are segmented liver slice,
which are overlaid with the original CT slice. It can be seen that accurate segmentations have been achieved.
Figure 5 (a) Histogram of the liver, (b) CT image, (c) intensity based probabilistic map.
Figure 6 Two typical segmentation results
We also compared our proposed method with the conventional method using ROC curves, which is shown in Figure 7. As shown in Figure 7, compared with the
conventional method (the red curve), our proposed method (the blue curve) can significantly improve both segmentation accuracy.
Figure 7 Comparison of our proposed method and the conventional method.
3.Conclusion
In this paper, we propose a template matching based framework for liver segmentation using probabilistic atlas. In our proposed method, we first find a Region of Interest (ROI) of the liver, which is based on human anatomic structure, and then the probabilistic atlas is used as a template to find the liver in the ROI by the use of template matching. Compared with the conventional method, our proposed method can significantly improve both segmentation accuracy and robustness. The proposed method can be extended to segment other organs.
Acknowledgments
This work is supported in part by the Grant-in Aid for Scientific Research from the Japanese Ministry for Education, Science, Culture and Sports (MEXT) under the Grant No. 2430076, No. 24103710, in part by the MEXT Support Program for the Strategic Research Foundation at Private Universities (2013-2017), and in part by the R-GIRO Research Fund from Ritsumeikan University.
References
[1]H.P. Meinzer, P.,Schemmer, M. Schobinger, M. Nolden, T. Heimann, B. Yalcin, G. M. Richte, T.Kraus, M. W. Buchler, M.Thorn, Computer-based Surgery Planning for Living Liver Donation. In: 20th ISPRS Congress, Istanbul 2004, International Ar- chives of Photogrammetry and Remote Sensing, vol.
XXXV(B), pp. 291–295, 2004.
[2]Y.-W. Chen, M. Kaibori, K. Miyawaki, T. Shindo, A. H. Foruzan, T. Tateyama, X.-H. Han, K. Matsui, T.
Tsuda, A-H. Kwon, Computer-Aided Liver Surgical Planning System Using CT Volumes, Proc. of IEEE EMBC2013, Osaka, Japan, pp.2360-2363, 2013.
[3]R. Adams, L Bischof, Seeded Region Growing, IEEE Trans. on Pattern Anal. Machine Intell., Vol. 16, No. 6, pp. 641-647, 1994.
[4]A. H. Foruzan, Y.-W. Chen at. al., “Segmentation of Liver in Low-contrast Images Using K-Means Clustering and Geodesic Active Contour Algorithms,” IEICE Trans., Vol.E96-D, pp.798-807, 2013.
[5] T. McInerney, D. Terzopoulos, Deformable Models in Medical Image Analysis: A Survey, Medical Image Analysis, Vol.1, No.2, pp.91-108, 1996
[6]F. Liu et al., “Liver segmentation for CT images using GVF snake,” Med.Phys. Vol.32, No.12, pp.3699–
3706. 2005.
[7]V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active contours,” Int. J. Comput. Vis. Vol.22, pp.61–
79, 1997.
[8] Y. Boykov, M. P. Jolly, Interactive organ segmentation using graph cuts. In: Delp, S.L., DiGoia, A.M., Jaramaz, B. (eds.) MICCAI 2000. LNCS, vol.1935, pp. 276–286. Springer, Heidelberg, 2000
[9]L. Massoptier, S. Casciaro, Fully Automatic Liver Segmentation through Graph-Cut Technique. In: Proceedings of the 29th Annual International Conference of the IEEE EMBS Cité Internationale, Lyon, France, August 23-26, 2007.
[10] Y.-W. Chen, K. Tsubokawa, and A. H. Foruzan, “Liver Segmentation from Low Contrast Open MR Scans Using K-Means Clustering and Graph-Cuts,” Lecture Notes in Computer Science, Springer, LNCS6064, pp.162-169, 2010.
[11] H. Park, P. Bland, C. Meyer, Construction of an abdominal probabilistic atlas and its application in segmentation, IEEE Transactions on Medical Imaging, Vol.22, pp.483–492, 2003.
[12] M. G. Linguraru, J. A. Sandberg, Z. Li, F. Shah, and R. M. Summers, Automated Segmentation and Quantification of Liver and Spleen from CT Images Using Normalized Probabilistic Atlases and Enhancement Estimation, Medical Physics, Vol.37, No.2, pp.771-783, 2009.
[13] T. Okada, R. Shimada, M. Hori, M. Nakamoto, Y.-W. Chen, H. Nakamaura, Y. Sato, Automated segmentation of the liver from 3D CT images using probabilistic atlas and multi-level statistical shape model, Academic Radiology, Vol.63, pp.1390-1403, 2008.
[14] X. Zhang, et al., Automatic liver segmentation from CT scans based on a statistical shape model, Proc.
of IEEE EMBC2010, pp.5351-5354, 2010.
[15] J. Staal, B. van Ginneken, M. A. Viergever, Automatic Rib Segmentation in CT Data, Lecture Notes in Computer Science Vol.3117, pp. 193-204, 2004.
[16] K. Zhao Kai, B. Kang, Y. Kang, H. Zhao, Auto-Threshold Bone Segmentation Based on CT Image and Its Application on CTA Bone-Subtraction, Symposium on Photonics and Optoelectronic (SOPO), 2010.
