Computerized Medical Imaging and Graphics
Volume 36, Issue 1 , Pages 25-37, January 2012

Left ventricular myocardium segmentation on arterial phase of multi-detector row computed tomography

  • I-Chen Tsai

      Affiliations

    • Department of Radiology, Taichung Veterans General Hospital, Taichung, Taiwan
    • Institute of Clinical Medicine, National Yang Ming University, Taiwan
    • ,
  • Yu-Len Huang

      Affiliations

    • Department of Computer Science, Tunghai University, Taichung 407, Taiwan
    • Corresponding Author InformationCorresponding author. Tel.: +886 4 23590121x33800; fax: +886 4 23591567.
    • ,
  • Kai-Hua Kuo

      Affiliations

    • Department of Computer Science, Tunghai University, Taichung 407, Taiwan

Received 25 November 2009; received in revised form 23 July 2010; accepted 18 March 2011. published online 15 April 2011.

Article Outline

Abstract 

Rationale and objectives

Variation of left ventricular myocardial volumes correlates closely with ischemic heart diseases. In clinical practice, because physicians and radiologists rely much on myocardial contour to diagnose many different cardiac diseases, automatic segmentation of left ventricular myocardium and quantifying myocardium characteristics is clinically beneficial. This paper presents a hybrid segmentation method for left ventricular myocardium on arterial phase of multi-detector row computed tomography (MDCT) imaging.

Materials and methods

The proposed method utilizes an intensity transformation equation as a preprocessing procedure to enhance contrast and reduce noise in MDCT imaging. By setting the centroid of left ventricle (LV) as an initial seed, the conventional region growing method is employed to identify the endocardial contour of LV cavity for each slice. Then the level-set method (LSM) utilizes the extracted endocardial contour as initial contour to delineate the epicardium of LV. The two extracted contours are integrated to form the region of interest (ROI) of the LV. Finally, the ROIs from all slices are combined to obtain the volume of the whole LV myocardium.

Results

Twenty-two healthy patients who had no symptoms of ischemic heart disease are applied to evaluate the performance of the proposed method. Compared with manual contours delineated by two experienced experts, the contouring results from computer simulation reveal that the proposed method always identifies similar contours as that obtained by the manual sketching.

Conclusion

The proposed method provides a robust and fast automatic contouring for LV myocardium on arterial phase of MDCT. The potential role of this technique may save much of the time required to manually sketch a precise contour with high stability.

Keywords: Multi-detector row computed tomography, Myocardium contouring, Level-set segmentation, Myocardial volume, Coronary heart disease

 

Back to Article Outline

1. Introduction 

Cardiac left ventricle (LV) receives oxygenated blood from the left atrium and pushes out blood into aorta for supplying human body with oxygen. Ischemic heart disease can reduce the cardiac output and cause heart failure or even death if severe. Thus, early diagnosing ischemic heart disease is very important. Both cardiac magnetic resonance (MR) and computed tomography (CT) imaging are currently the most widely used screening modalities for diagnosing LV global or regional wall motion problems [1]. Generally radiological researchers believe that MR has higher temporal resolution and thus might provide more accurate functional data than CT. However, multi-detector row computed tomography (MDCT) has the advantage in coronary image and is widely used in many hospitals nowadays. Cardiac MDCT has been shown as a promising tool in the diagnosis of ischemic heart disease [2]. Physician and radiologist could observe the condition of left ventricular myocardium to make diagnostic decision. The myocardial variation on arterial phase of MDCT is clinically significant due to thinning of the myocardium of LV probably means myocardial infarction. Measuring myocardium volumes on both arterial and delayed phases of MDCT images can help in determining the infarction volume of LV [3].

However, manual myocardial contouring on all MDCT slices is time-consuming and tiring [4]. Repetitive operation of delineating LV myocardial contour and distinguishing contour variation by experts gives rise to increase of delineating faults. Not all slices of LV are distinct, some are noisy and hard to delineate contour of LV myocardium immediately. Observer must experimentally change the contrast and intensity of images to inspect a clarity boundary of myocardium. An automatic segmentation for left ventricular myocardium in MDCT imaging could provide a swift way to quantify the myocardial volumes. Moreover, a precise segmentation supports serviceable LV myocardial analyses including functional assessment (e.g. ejection fraction [5] and regional wall motion [6]) and physiological assessment (e.g. viability [7], [8], [9], myocardial thickness [10] and perfusion [11]). As MDCT imaging for diagnosis of ischemic heart disease becomes increasingly widespread, the clinical application of an effective automatic contouring method is becoming urgent. Several LV myocardium segmentation methods for cardiac MR imaging have been developed [12], [13], [14], [15] and demonstrate robust performance. However, by comparing MDCT and MR imaging, MDCT has higher spatial resolution, lower contrast resolution and great CT number quantification, these fundamental differences make the segmentation approaches for these two modalities significantly different. An automatic segmentation of the LV myocardium on both MR and CT images was first introduced by Jolly [16]. However, the segmentation is performed on each slice individually, without exploiting the relationship between the adjacent slices. Due to the contours of LV myocardium on serial short-axis MDCT slices tend to be highly correlated, this study developed an auto-segmentation method by exploiting the correlation between inter-slices to reduce analysis time and delineate contour consistently and correctly.

The conventional edge-based segmentation methods often employ the gradient of the image to identify regions’ boundary. These methods are not designed for detecting the discontinuity of image intensity. Thus the edge-based methods do not perform well for the MDCT image segmentation. The region-based segmentation methods such as split-and-merge, region-growing [17], snake deformation [18], [19] and morphological watershed transformation [20] are sensitive to the noise and contrast in an image. These methods are also unsatisfied for extracting myocardium because of the different anatomical structure of endocardium and epicardium. A number of studies combined the segmentation methods with a shape model gathered statistics from training data set to improve segmentation accuracy [21], [22], [23]. However, constructing a good model is strenuous and requires sufficient samples for the training procedure.

Instead of segmenting LV myocardium directly on arterial phase MDCT imaging, the proposed method employed a hybrid technique which delineated endocardial contour and epicardial contour separately. The proposed method comprised three phases, i.e. preprocessing, endocardial contouring and epicardial contouring. The preprocessing phase utilized a predictable intensity transformation [24] which overcame the left ventricular contour inconspicuous to improved accuracy of segmentation substantially. The endocardial contouring phase identified the cavity of LV for delineating endocardial contour by using the region growing method [17]. When region growing method finished, determining the endocardial contour was still a complicated task due to papillary muscle is similar and close to myocardium. In this work, the proposed method performed convex hull algorithm to remove the influence of papillary muscle. In the epicardial contouring phase, a deformable contour method, the threshold level-set segmentation method (LSM) [25], [26] exploited the extracted endocardial contour as initial contour to delineate precise epicardial contour. The LSM segmentation conquered the difficulty of distinguishing left ventricular myocardium from right ventricular myocardium and other adjacent tissues. Eventually, the left ventricular myocardium could be represented as the region of interest (ROI) area between the two extracted contours and then the integrated information of myocardium could be obtained for further analysis.

Back to Article Outline

2. Materials and methods 

2.1. Image acquisition 

Unhealthy patient heart with pseudoaneurysm or myocardial thinning is occupying only less than 5% of cardiac scan in our clinical practice. We decided to develop this approach based on the 95% healthy left ventricle first. In this study, the database of the MDCT images was from 22 healthy volunteers (from 40 to 80 years old). The volunteers had no hypertension, smoking history, family history or symptoms of ischemic heart disease or history of cardiomyopathy. The cardiac CT scans were performed by using a 40-detector row CT (Brilliance 40; Philips, Best, the Netherlands) and a senior CT technologist with more than 10 years of experience performed all of the scans. Each volunteer was injected contrast agent. Next, arterial phase scan was applied a bolus-tracking technique, which started 7-s delayed scan when ascending aorta met the threshold of 150 HU. All parameters for MDCT scan were described in detail in our previous work [3]. Fig. 1 illustrates that serial short axis images covering the entire LV had been created from apex of the heart. Thickness for short axis images was 5mm. The first five slices counting from the bottom are defined as apex (Apex), the sixth to tenth slices are defined as mid-ventricular level (MidV) and the remaining slices are defined as base (Base). All the data are stored as DICOM format with capturing resolution 512×512 pixels. The monochrome image is quantized into 8 bits (i.e. 256 gray levels).

2.2. The proposed segmentation method 

Each MDCT scan from apex to base obtained approximately 20 serial slices of LV. Any two adjacent slices are practically to be similar, thus the proposed method exploited the relationship between the adjacent slices to segment the LV myocardial contours. Due to the myocardial boundaries in the MidV slices were always regular and obvious, the slice at middle-ventricle was selected as the beginning slice in the proposed contouring procedure. Fig. 2 shows an example of the formed ROI in the middle-ventricle slice by combining the extracted endocardial and epicardial contours. Let Si is the slice i in an MDCT imaging. After extracting the contours from Si, the obtained contours were then utilized as the initial contours for segmenting the corresponding preceding slice Si−1 or the following slice Si+1. When the contour was disappeared or the extracted ROI was unreasonable, the iterative contouring procedure would be terminated. After all contours in MDCT slices were acquired, the three-dimension (3D) contour of left ventricular myocardium would be constructed and the volume could be estimated. Fig. 3 presents the flowchart of the proposed method which includes the preprocessing, endocardial contouring and epicardial contouring phases.

  • View full-size image.
  • Fig. 2. 

    The proposed segmentation for the middle-ventricle slice on an MDCT imaging: (a) original image, (b) preprocessed image, (c) result of the region growing process, (d) the extracted endocardial contour, (e) the extracted epicardial contour and (f) the formed ROI.

2.2.1. Preprocessing 

Due to the MDCT images of LV always contain noises, this study performed an effective de-nosing filter to diminish the influence of noise for the following contouring phases. Traditional de-nosing filters constantly blur the sharp boundary of anatomical structures constantly. Perona and Malik [27] proposed a linear-filter, i.e. the anisotropic diffusion filter, to reduce noise with a high performance. The anisotropic diffusion filtering is not only distorting the fine structure on Gaussian noisy images but also efficiently reserving the anatomical boundaries. The proposed method employed the anisotropic diffusion filter to smooth myocardium of LV and to weaken the speck caused by papillary muscle. The anisotropic diffusion solution is given by:

(1)
where g(x, y, t) is the image result of every time slice, x and y are the coordinates of two dimension, and t is time parameter. While t=0, g(x, y, 0) is the initial image. The equation of anisotropic diffusion includes a variable, the conductance term, which can prevent edges from being smoothed:
(2)
where k is the parameter controlling the amount of diffusion near object edges. This study performs the filter with k=5 to enhance the MDCT images. Fig. 4(a) illustrates the de-nosed result by applying the filter to the original image (Fig. 2(a)), Fig. 4(b) is the difference image between Fig. 2, Fig. 4. Obviously, the anisotropic diffusion method effectively reduced the mass of noises without blurring the boundary on myocardium.

This study also applied an intensity transformation to adjust the intensity value of the de-noised images. The intensity transformation was used to enhance image contrast before segmentation procedure. In order to make the endocardial contouring effortless, a common intensity transformation was utilized to obtain dissimilarity between LV cavity and myocardium as much as possible. The transformation is formed as s=T(r), where variables r and s are the intensity of pixel at location (x, y) of input image and output image, respectively. The transformation raised intensity of the LV cavity to the highest value but maintained the intensity of myocardium. In other words, T(r) increases the input intensity if r is higher than the factor α, representing an assumed intensity of LV myocardium; otherwise T(r) decreases intensity if r is lower than α. T(r) is represented as

(3)
where r and α were normalized into [0,1]. The factors α and β were distinct in each image and the factor β affects the amplitude of curve. The larger value of β made larger adjustment of intensity r, the default value of factor β was set for 1.

In this study, the factor α was determined based on intensity histogram of the de-noised image. The brightest region on histogram was estimated by 10% pixels with the highest intensity in an image. An intensity median γ of the brightest region on histogram was firstly picked as an assumed intensity of LV cavity, which taken the region of the brightest peak while contrast agent enters LV with blood stream. Then the median γ was utilized to decide α by the ratio of LV myocardial area to cavity area. With the ratio set for 1.5, i.e. α=2/3×γ, the proposed contouring scheme obtained a stable and the highest accuracy. Fig. 5 illustrates an example of intensity transformation. The LV cavity in Fig. 5(a) was transformed into a white region (see Fig. 5(b)) and the boundaries between myocardium and other tissues were distinctly enhanced.

2.2.2. Endocardial contouring 

The endocardial contouring phase included automatic seed searching, region growing and post-processing procedures. The seed searching procedure located a proper pixel within the LV cavity as the seed for region growing segmentation. A thresholding step was utilized to identify the biggest region near the center of image. The threshold τ was determined as the average of the enhanced intensity of LV cavity, this study set τ=(1+α)/2. The centroid of the detected region was then regarded as the seed. Fig. 6 shows the process of the automatic seed searching.

  • View full-size image.
  • Fig. 6. 

    Automatic seed searching: (a) binary image that generated by thresholding on the enhanced image, (b) the speckle removed image (using the erosion operator) and (c) located seed from the biggest region near the centroid of image.

The threshold τ was also applied in the region growing procedure. If the intensity value of a pixel is higher than τ, the pixel was probably to be inside the cavity of LV. The region growing method gathers pixels together by using a simple criterion. The method is based on intensity information and chooses adjacent pixels which have similar intensity. The criterion is given by . Scope of lower and upper was set to τ and 1, respectively. The LV cavity region including seed was segmented beyond the scope of lower and upper.

The drawback of region growing segmentation is that the intensity provides minor information for distinguishing the similar tissues. Fig. 2(b) and (c) shows that there are many speckles near the cavity wall of LV. The speckles were generated by papillary muscle with the same intensity with LV myocardium and caused the region growing result would not locate on the actual boundary of LV cavity. While papillary muscle made a large speckle or was joined to myocardium, region growing method produced an unsuccessful segmentation. Therefore the Quickhull algorithm [28], which is considered to be a practical convex hull finding method for polygons, was performed as the contour post-processing to fill pits on the extracted boundary (see Fig. 2(c)). The convex hull of a set of points is the smallest convex set that contains the boundary points obtained by the region growing segmentation. By comparing Fig. 7(a) with (b), the convex hull algorithm improved the accuracy of the endocardial contouring.

  • View full-size image.
  • Fig. 7. 

    A comparison of the endocardial and epicardial contours: (a) the endocardial contour generated by region growing and (b) the improved contours using the convex hull algorithm.

2.2.3. Epicardial contouring 

Due to the intensity of right ventricular myocardium is similar to some tissues connected to LV myocardium, the region growing method is inappropriate for the epicardial contouring (see Fig. 8(a)). This study performed a sophisticated segmentation algorithm, level-set method (LSM), to achieve epicardial contouring. The LSM, a useful deformable segmentation method, utilizes the initial contour for iteratively tracking contour. The proposed utilized the extracted endocardial contour as initial contour for the LSM segmentation to obtain LV epicardial contour.

The LSM computing and analyzing the curve propagation has been successfully used for solving complicated segmentation problem on images. The concept of LSM is to embed contour as the zero level-set of a higher dimensional function called the level-set function Ψ(x, t). Computing zero level-set, Γ={(x, t)|Ψ(x, t)=0}, obtains evolving contour at any time t. A generic level-set equation is given by [29]

(4)
where A is an advection term, P term is a propagation term, and Z is a spatial modifier term for the mean curvature κ. The constants a, b, c are the weight for A, P, Z terms, separately. However, LSM probably caused epicardial contour to be ‘leak’ (as shown in Fig. 8(b)). A threshold level-set method [30] improved P term which utilized the intensity information to change the propagation direction. The P term was computed in intensity from an upper threshold U and a lower threshold L of input image g
(5)

Fig. 9 indicates the relation of intensity to propagation term. While the intensity value g(x) is inside of range of U and L, P has a positive value to make contour expand. Otherwise, contour contrasts with a negative P. Let α′ be the average of U and L, and bounded range is the same as (1−α). The upper threshold and lower threshold are defined as

(6)
where α′ is the mean intensity of pixels along the initial contour. The mean intensity α′ was more accurate than α experimentally. According to observation on intensity transformation curve, the bright image has narrow range from U to L because (1−α) is small; oppositely, dark image has wide range from U to L. Hence, epicardial contour was prevented from under-segmentation (dark images) and over-segmentation (bright images).

The segmentation result of threshold level-set method required a final processing procedure to refine the contour. The morphological operator closing was applied to fill contour pits caused by the speckles of LV cavity. The operator opening was applied to eliminate contour peaks caused by right ventricular myocardium. Fig. 2(e) shows the result of epicardial contouring phase and Fig. 2(f) illustrates the combination of endocardial contour and epicardial contour.

2.2.4. Contour validation 

In this study, the extracted contours were provided to the segmentation of the next image slice as referencing information. An inconsistent ROI would obtain an inaccurate segmentation of the next image slice. In order to prevent an inaccurate segmentation, the automatic contouring procedure would be terminated while an inconsistent ROI was acquired. The proposed method performed a contour validation procedure on each MDCT slice. The procedure compared the current extracted ROI with previous one in seed location, ROI area and extension scope of epicardial contour.

Generally, the location of seed should be on the inside of an LV cavity. An illegal seed would cause an unsuccessful region growing while the seed was located on ROI or outside ROI. The validity of seed was checked by the threshold τ. If intensity of seed was smaller than τ, the seed was considered as an illegal seed and then a new seed was determined from measuring the centroid of adjacent endocardial contour. Moreover, ROI area variation between two adjacent slices would slight (under 20% variation clinically). Thus the ROI area variation over 30% always indicates the occurrence of over-segmentation or under-segmentation. Due to a left ventricular outflow tract at the Apex or Base slices might obtain an unsuccessful segmentation, the validation was performed on checking the boundary of epicardial contour. When the extracted contour was considered as an over-segmentation or under-segmentation contour, the contouring procedure for the specific sequence (forward/backward) would be terminated.

2.3. Contour evaluation 

This study performed practical similarity measures to evaluate the quality of results in a quantitative and objective way. Fig. 10 demonstrates that contour segmented by the proposed method (SEG) and contour segmented by experts (REF) are overlapped. The term REF∩SEG denotes the area of overlap. The areas of miss and extra represent segmentation errors of the proposed method. The four measures assessed the similarity between REG area and SEG area [31], i.e. similarity index (SI), overlap fraction (OF), extra fraction (EF) and overlap value (OV). The definitions of the indices are

(7)
(8)
(9)
(16)

  • View full-size image.
  • Fig. 10. 

    Comparison of an automated contouring area (SEG) with the manual contouring area (REF), with (overlap) the correctly segmented pixels (extra) the false positives and (miss) the false negatives.

When SI, OF, and OV are close to 1, and EF computation is close to 0, it means that the contours generated by automatic segmentation is similar to the manual contours.

Back to Article Outline

3. Results 

In this study, the dataset was acquired form 22 patients on arterial phase of MDCT, and finally 346 contours were obtained. Two experts (denoted Expert#1 and Expert#2) manually delineated contours of left ventricular myocardium individually. Expert#1 is a senior technologist with six years of experience on cardiac MDCT, Expert#2 is an experienced senior cardiac radiologist with six years experience. The four similarity measures between the manually determined contours and the automatically detected contours were evaluated.

Fig. 11 shows the result applied the proposed method on a patient's MDCT imaging. Fig. 12 illustrates the contours delineated by one of the experts that corresponded with Fig. 11. The contouring results of the proposed method were similar to that of the experts. The automatic segmentation eliminated the influences of papillary muscle on endocardial contour and smoothed undesired peaks on epicardial contour, even in images with distinct intensity and contrast. Table 1, Table 2 list the average indices of the proposed automatic endocardium and epicardium contouring. Table 3 shows the assessment indices between two expert's manual contours. The proposed method clearly yields contour that are similarly to those manually sketched. Only a small number of slices (less than 5%) in a case might generate an undesired segmentation in the automatic contours.

Table 1. The contouring evaluations of the proposed method which compared with Expert#1 using the measurements.
Case#EndocardiumEpicardium
SIOFEFOVSIOFEFOV
10.9500.9330.0320.9050.9440.9340.0450.894
20.9330.9140.0450.8750.9390.9170.0360.886
30.9150.8990.0610.8460.9230.8760.0120.865
40.9100.8430.0090.8350.9550.9600.0520.914
50.9160.8510.0070.8460.9560.9370.0220.917
60.9130.8410.0010.8410.9560.9680.0590.916
70.9200.8640.0140.8530.9600.9420.0210.923
80.9540.9190.0080.9110.9720.9690.0260.945
90.9060.8440.0180.8290.9400.9030.0150.889
100.9020.8230.0010.8230.9590.9470.0260.923
110.9320.9350.0970.8820.9600.9680.0540.925
120.9230.8630.0060.8580.9640.9520.0240.930
130.9260.8640.0010.8630.9660.9590.0270.934
140.9190.8530.0010.8520.9620.9470.0210.928
150.9210.8560.0020.8540.9650.9660.0360.932
160.9200.9020.0580.8550.9230.8700.0110.860
170.9430.9150.0250.8930.9530.9230.0120.912
180.9380.9170.0370.8840.9550.9460.0360.913
190.9020.8620.0640.8250.9610.9580.0380.925
200.9540.9300.0190.9120.9650.9530.0220.933
210.9420.8960.0060.8910.9730.9620.0170.947
220.9320.8820.0110.8720.9460.9010.0050.897
Average0.9260.8820.0240.8640.9540.9390.0280.914

SI: similarity index; OF: overlap fraction; EF: extra fraction; OV: overlap value.

Table 2. The contouring evaluations of the proposed method which compared with Expert#2 using the measurements.
Case#EndocardiumEpicardium
SIOFEFOVSIOFEFOV
10.9360.8860.0060.8810.9440.9340.0450.894
20.9190.8550.0050.8510.9620.9570.0330.927
30.9150.8580.0120.8480.9380.9090.0170.893
40.8920.8060.0000.8060.9630.9560.0300.928
50.9080.8340.0010.8330.9660.9560.0240.934
60.8970.8140.0000.8140.9660.9760.0450.934
70.9170.8470.0010.8460.9660.9520.0190.934
80.9400.8900.0030.8870.9730.9680.0220.947
90.9280.8780.0140.8660.9360.8940.0120.883
100.9080.8330.0000.8320.9600.9420.0200.924
110.9330.9140.0590.8800.9610.9580.0390.926
120.9220.8580.0010.8570.9650.9450.0130.933
130.9370.8830.0010.8820.9700.9600.0200.941
140.9370.8870.0060.8820.9650.9450.0120.933
150.9300.8710.0010.8690.9700.9630.0220.942
160.9150.8630.0210.8450.9250.8680.0060.863
170.9540.9240.0130.9120.9520.9140.0050.910
180.9320.8820.0110.8730.9610.9650.0450.925
190.9240.8700.0120.8600.9650.9490.0180.932
200.9480.9090.0080.9010.9680.9590.0220.938
210.9430.8950.0040.8920.9680.9470.0090.939
220.9250.8680.0080.8610.9440.9020.0090.894
Average0.9250.8690.0090.8630.9580.9420.0220.922

SI: similarity index; OF: overlap fraction; EF: extra fraction; OV: overlap value.

Table 3. The contouring evaluations of Expert#2 (set as SEG) which compared with Expert#1 (set as REF).
Case#EndocardiumEpicardium
SIOFEFOVSIOFEFOV
10.9470.9110.0140.8990.9630.9710.0460.929
20.9390.8920.0080.8850.9620.9800.0580.927
30.9400.8990.0120.8880.9620.9820.0610.927
40.9460.9220.0260.8990.9670.9560.0200.937
50.9600.9480.0270.9230.9710.9810.0400.943
60.9620.9460.0210.9270.9700.9680.0270.943
70.9590.9440.0240.9220.9750.9790.0300.950
80.9690.9530.0130.9400.9820.9790.0160.964
90.9530.9700.0670.9100.9680.9620.0250.939
100.9670.9730.0390.9370.9760.9700.0180.954
110.9620.9400.0140.9270.9720.9610.0150.946
120.9610.9560.0320.9260.9740.9650.0160.949
130.9730.9830.0390.9470.9790.9750.0180.958
140.9590.9810.0670.9210.9730.9670.0210.947
150.9670.9750.0410.9370.9740.9650.0180.949
160.9420.9130.0220.8930.9730.9660.0200.947
170.9670.9660.0320.9350.9730.9650.0190.948
180.9460.9190.0210.8990.9670.9810.0480.936
190.9510.9460.0360.9120.9720.9600.0140.947
200.9640.9490.0200.9300.9790.9830.0250.959
210.9650.9640.0340.9320.9740.9630.0140.950
220.9620.9540.0280.9280.9780.9800.0250.956
Average0.9570.9460.0290.9190.9720.9710.0270.946

SI: similarity index; OF: overlap fraction; EF: extra fraction; OV: overlap value.

Moreover, this study made a comparison for area of myocardium (mm2) between the automatic and manual segmentations. Fig. 13(a) shows scatter plot of the proposed automatic segmentation against the mean of manual segmentations from two experts for all datasets. The result revealed the proposed method could provide a stable and swift procedure to quantify the LV myocardial volumes. Fig. 13(b) shows the cumulative distributions of error distances. The proposed method made average distance error of 3.7 pixels. And for 95% of the contour points, the distance error is less than 10 pixels. All analyses were made on a single CPU Intel Pentium-VI 3.0GHz personal computer (ASUSTek Computer Inc., Taipei, Taiwan) with Microsoft Windows XP operating system. The programs were performed using Visual Studio C# 2005, Insight toolkit (ITK) and Matlab software (The MathWorks, Inc., Natick, MA). Average execution time of each case was less than 30s.

  • View full-size image.
  • Fig. 13. 

    (a) Scatter plot of the automatic segmentation (SEG) against the mean of manual segmentation from two experts (REF) for all datasets and (b) cumulative distribution of error distance between manual contours and the automatically segmented contours.

Back to Article Outline

4. Discussion and conclusion 

The proposed method automatically generated the endocardial and the epicardial contours for each MDCT slice in sequence. The accuracy of endocardial contour influenced greatly the epicardial contouring, as shown in Fig. 7(a). Suppose that an endocardial contour was restricted by speckles and then generated an unsatisfied initial contour, a formation of epicardial contour would take more inward traction that making contour stay with endocardial wall (short arrow in Fig. 7(a)). The inward traction also resulted in under-extension contour at hypertrophy place of myocardium (long arrow in Fig. 6(a)).

The concaves appeared at the Base slices that always caused by the papillary muscle and left ventricular outflow tract would decrease assessment averages. The proposed method utilized the convex hull algorithm to avoid concaving and than to obtain polished contours. However, an unsatisfied segmentation would be obtained when the concave is more conspicuous than enough. Fig. 14(a) shows an unfavorable contour generated by the proposed method and Fig. 14(b) is the manual sketched contour. The indices (SI, OF, EF, OV) of the automatic endocardium and epicardium contouring were (0.844, 0.887, 0.214, 0.732) and (0.857, 0.883, 0.181, 0.759), respectively.

In this study, the effect of papillary muscle was reduced by using region growing method. The deformable contouring methods such as the LSM might acquire an inaccurate contour which did not reach boundaries of small gaps. Although parameters of the LSM could be modified for enhancing the segmentation ability on small pits, but breaking the smooth contour is highly probable. Fig. 15 illustrates the situation, the contouring results of the LSM and the region growing methods, respectively.

Automatic segmentation for LV myocardium may assist radiologists and technologists, without relevant experience, in quantifying the infarct size and predicting the patient's prognosis. For this purpose, this study proposed an automatic segmentation of LV on arterial phase of MDCT imaging. The proposed method could reliably delineate the contours of left ventricular myocardium. The segmentation result compared with manual contours was 94% in term of similarity index. From the simulations, the proposed method would be performed to evaluate the morphological features about heart functions. This study is not to emphasize that the automatic contouring technique is superior to the one undertaken manually. In our opinion, both automatic and manual contours were not necessarily same factual border after all. However, automatic segmentation can save much of the time required to sketch precise contours with very high stability. In the future, the applicability of this approach to diseased heart will be checked. If using both healthy and unhealthy heart scan to develop the model, due to the rare and extreme heart condition (e.g. pseudoaneurysms, LV rupture, myocardial thinning, etc.), the model might be biased and not useful in the 95% healthy hearts. Moreover, we will try to automatically segment the delayed phase of MDCT and perform the techniques on unhealthy hearts. The purpose of this work can be used to analyze diseases of LV from patients by classification and statistical methods. The development of the automatic contouring method is indeed important and its medical application is urgent.

Back to Article Outline

Acknowledgments 

The authors would like to thank the Taichung Veterans General Hospital (Taiwan) and Tunghai University for financially supporting this research under Contract No. TCVGH-T987811. This research was supported in part by National Science Council of the Republic of China (Taiwan) under Contract No. NSC98-2221-E-029-026.

Back to Article Outline

References 

  1. Dewey M, Muller M, Eddicks S, Schnapauff D, Teige F, Rutsch W, et al. Evaluation of global and regional left ventricular function with 16-slice computed tomography, biplane cineventriculography, and two-dimensional transthoracic echocardiography: comparison with magnetic resonance imaging. J Am Coll Cardiol. 2006;48:2034–2044
  2. Tsai IC, Lee WL, Tsao CR, Chang Y, Chen MC, Lee T, et al. Comprehensive evaluation of ischemic heart disease using MDCT. AJR Am J Roentgenol. 2008;191:64–72
  3. Tsai IC, Huang YL, Liao WC, Kuo KH, Chen MC. Left ventricular myocardial volumes measured during arterial and delayed phases of multidetector row computed tomography: a study on intra- and interobserver variability. Int J Cardiovasc Imaging (formerly Cardiac Imaging). 2009;25:55–63
  4. Boehm T, Alkadhi H, Roffi M, Willmann JK, Desbiolles LM, Marincek B, et al. Time-effectiveness, observer-dependence, and accuracy of measurements of left ventricular ejection fraction using 4-channel MDCT. Rofo. 2004;176:529–537
  5. Okuyama T, Ehara S, Shirai N, Sugioka K, Ogawa K, Oe H, et al. Usefulness of three-dimensional automated quantification of left ventricular mass, volume, and function by 64-slice computed tomography. J Cardiol. 2008;52:276–284
  6. Mahnken AH, Katoh M, Bruners P, Spuentrup E, Wildberger JE, Gunther RW, et al. Acute myocardial infarction: assessment of left ventricular function with 16-detector row spiral CT versus MR imaging – study in pigs. Radiology. 2005;236:112–117
  7. Mahnken AH, Koos R, Katoh M, Wildberger JE, Spuentrup E, Buecker A, et al. Assessment of myocardial viability in reperfused acute myocardial infarction using 16-slice computed tomography in comparison to magnetic resonance imaging. J Am Coll Cardiol. 2005;45:2042–2047
  8. Sato A, Hiroe M, Nozato T, Hikita H, Ito Y, Ohigashi H, et al. Early validation study of 64-slice multidetector computed tomography for the assessment of myocardial viability and the prediction of left ventricular remodelling after acute myocardial infarction. Eur Heart J. 2008;29:490–498
  9. Lardo AC, Cordeiro MA, Silva C, Amado LC, George RT, Saliaris AP, et al. Contrast-enhanced multidetector computed tomography viability imaging after myocardial infarction: characterization of myocyte death, microvascular obstruction, and chronic scar. Circulation. 2006;113:394–404
  10. Koyama Y, Matsuoka H, Mochizuki T, Higashino H, Kawakami H, Nakata S, et al. Assessment of reperfused acute myocardial infarction with two-phase contrast-enhanced helical CT: prediction of left ventricular function and wall thickness. Radiology. 2005;235:804–811
  11. George RT, Jerosch-Herold M, Silva C, Kitagawa K, Bluemke DA, Lima JA, et al. Quantification of myocardial perfusion using dynamic 64-detector computed tomography. Invest Radiol. 2007;42:815–822
  12. Lee HY, Codella N, Cham M, Prince M, Weinsaft J, Wang Y. Left ventricle segmentation using graph searching on intensity and gradient and a priori knowledge (lvGIGA) for short-axis cardiac magnetic resonance imaging. J Magn Reson Imaging. 2008;28:1393–1401
  13. Lynch M, Ghita O, Whelan PF. Left-ventricle myocardium segmentation using a coupled level-set with a priori knowledge. Comput Med Imaging Graph. 2006;30:255–262
  14. Lynch M, Ghita O, Whelan PF. Segmentation of the left ventricle of the heart in 3-D+t MRI data using an optimized nonrigid temporal model. IEEE Trans Med Imaging. 2008;27:195–203
  15. Codella NC, Weinsaft JW, Cham MD, Janik M, Prince MR, Wang Y. Left ventricle: automated segmentation by using myocardial effusion threshold reduction and intravoxel computation at MR imaging. Radiology. 2008;248:1004–1012
  16. Jolly MP. Automatic segmentation of the left ventricle in cardiac MR and CT images. Int J Comput Vision. 2006;70:151–163
  17. Adams R, Bischof L. Seeded Region Growing. IEEE Trans Pattern Anal Mach Intell. 1994;16:641–647
  18. Dagher I, El Tom K. WaterBalloons: a hybrid watershed Balloon Snake segmentation. Image Vision Comput. 2008;26:905–912
  19. Kass M, Witkin A. Terzopoulos D: snakes – active contour models. Int J Comput Vision. 1987;1:321–331
  20. Vincent L, Soille P. Watersheds in digital spaces – an efficient algorithm based on immersion simulations. IEEE Trans Pattern Anal Mach Intell. 1991;13:583–598
  21. Lorenzo-Valdes M, Sanchez-Ortiz GI, Elkington AG, Mohiaddin RH, Rueckert D. Segmentation of 4D cardiac MR images using a probabilistic atlas and the EM algorithm. Med Image Anal. 2004;8:255–265
  22. Lotjonen J, Kivisto S, Koikkalainen J, Smutek D, Lauerma K. Statistical shape model of atria, ventricles and epicardium from short- and long-axis MR images. Med Image Anal. 2004;8:371–386
  23. Hamarneh G, Li X. Watershed segmentation using prior shape and appearance knowledge. Image Vision Comput. 2009;27:59–68
  24. Gonzalez RC, Woods RE. Digital image processing. 3rd ed.. New Jersey: Pearson Prentice Hall; 2010;
  25. Huang YL, Jiang YR, Chen DR, Moon WK. Level set contouring for breast tumor in sonography. J Digit Imaging. 2007;20:238–247
  26. Malladi R, Sethian JA, Vemuri BC. Shape modeling with front propagation – a level set approach. IEEE Trans Pattern Anal Mach Intell. 1995;17:158–175
  27. Perona P, Malik J. Scale-space and edge-detection using anisotropic diffusion. IEEE Trans Pattern Anal Mach Intell. 1990;12:629–639
  28. Barber CB, Dobkin DP, Huhdanpaa HT. The Quickhull algorithm for convex hulls. ACM Trans Math Software. 1996;22:469–483
  29. Ibanez L, Schroeder W, Ng L, Cate J. The ITK guide – updated for ITK version 2.4, 2nd ed.; available online at http://www.itk.org/ItkSoftwareGuide.pdf, Nov. 2005.
  30. Lefohn AE, Cates JE, Whitaker RT. Interactive GPU-based level sets for 3D segmentation. Med Image Comput Comput Assist Interv – Miccai. 2003;2878:564–572
  31. Anbeek P, Vincken KL, van Osch MJ, Bisschops RH, van der Grond J. Probabilistic segmentation of white matter lesions in MR imaging. Neuroimage. 2004;21:1037–1044

PII: S0895-6111(11)00046-2

doi:10.1016/j.compmedimag.2011.03.003

Computerized Medical Imaging and Graphics
Volume 36, Issue 1 , Pages 25-37, January 2012

Article Tools

* Image Usage & Resolution