Medical & Biological Engineering & Computing

, Volume 45, Issue 11, pp 1143–1152

Robust border enhancement and detection for measurement of fetal nuchal translucency in ultrasound images

Authors

  • Yu-Bu Lee
    • Department of Computer Science and EngineeringEwha Womans University
  • Min-Jeong Kim
    • Department of Computer Science and EngineeringEwha Womans University
    • Department of Computer Science and EngineeringEwha Womans University
    • Center for Computer Graphics and Virtual RealityEwha Womans University
Special Issue

DOI: 10.1007/s11517-007-0225-7

Cite this article as:
Lee, Y., Kim, M. & Kim, M. Med Bio Eng Comput (2007) 45: 1143. doi:10.1007/s11517-007-0225-7

Abstract

Ultrasonic measurement of nuchal translucency (NT) thickness in the first trimester of pregnancy has recently been proposed as the most useful marker in early screening for fetal chromosomal abnormalities. However, manual tracing of the two echogenic lines in the image, using on-screen calipers, is hampered by weak edges, together with noise and other artifacts, leading to variable results and inefficiency. Our semi-automatic method of fetal NT thickness measurement uses a coherence-enhancing diffusion filter to enhance the border and reduce noise, followed by detection of the NT by minimization of a cost function, that combines intensity, edge strength and continuity, using dynamic programming. This method has been validated by determining the correlation between manual and semi-automatic measurements.

Keywords

Fetal nuchal translucencyEdge-enhancing filteringBorder detectionDynamic programmingSemi-automatic measurement

1 Introduction

Ultrasonography is performed during early pregnancy for dating, determination of the number of fetuses, assessment of early complications, and increasingly for evaluation of the fetus, including measurement of the nuchal translucency (NT) thickness [7]. Measurement of NT thickness in the first trimester of pregnancy has proved to be one of the most discriminating prenatal markers in screening for chromosomal abnormalities such as trisomies 13, 18 and 21 [14]. Furthermore, an increased NT thickness (>2.5 mm) between 10 and 14 weeks has also been associated with an increased risk of congenital heart and genetic syndromes [9, 15, 19].

The NT thickness is measured in the sagittal section of the fetus by transabdominal or transvaginal ultrasound examination. To determine the NT thickness, most physicians trace the boundaries of the NT layer manually, but this is time-consuming, and the results vary according to the training, experience, and subjective judgment of the observer.

Ultrasound images frequently include weak echoes, echo dropouts, and speckle noise, which make feature extraction, analysis and quantitative measurement difficult and unreliable. Therefore, techniques for the suppression of noise and for feature enhancement have been developed to improve the accuracy and reliability of segmentation. Commonly used linear low-pass filters, such as the mean filter and Gaussian filter, are not suitable for reducing the speckle noise of ultrasound images since they eliminate the high frequencies and thus tend to smooth out edges in the image [6]. The median filter is arguably the most popular nonlinear filter, and is well suited to the reduction of speckle-type structures, but it too has limited success in preserving structural features such as the weak, diffuse edges that characterize ultrasound images. Perona and Malik [13] have proposed anisotropic filtering, based on the diffusion theory, which is better able to distinguish regional structures, but can sometimes blur thin structures. It is possible to enhance thin edges by considering the local coherence of structures, and Weickert [16] has proposed a coherence-enhancing diffusion (CED) filter, which reduces speckle noise using a structure tensor, while improving the image contrast.

To avoid the problems inherent in manual measurements, many automated techniques have been proposed. Bernardino et al. [4] developed a semi-automatic measurement system, which uses the Sobel operator to detect the border of the NT layer. However, well-known edge-detection techniques such as the Sobel operator or the Canny’s method determine the location of an edge by local evaluation of a single image feature such as intensity or the intensity gradient. But no single image feature can provide reliable border measurement in fetal ultrasound images.

In their work on ultrasonic images of the carotid artery, Cheng et al. [5] suggested an automatic system for detecting the boundaries of the intima-media complex based on the snake. The snake is very sensitive to the initial contour. If the initial contour is placed far away from the boundary of interest then the snake will not be attracted. The method still has limitation that the starting and ending points of the initial contour must be defined by user interaction. Loizou et al. [11] proposed a snake segmentation method, which utilizes normalization and speckle reduction as preprocessing. This method proposed an initialization procedure for improving snake initialization. However, the procedure needs many steps such as ROI definition, filtering, Otsu’s thresholding, dilation, labeling connecting components, etc., a discussion about irregular boundaries is missing. A computerized method to automatically detect the boundaries of intima-media using dynamic programming (DP) was proposed [8, 10, 17], which takes multiple-image features into account. By considering multiple-image features, such as intensity, gradient and border continuity, this method has obtained accurate and reproducible automated ultrasonic measurements. However, the system did not consider speckle-reduction filtering for dealing with poor images, which have a high level of speckle noise and weak edges and also require some manual correction.

In this paper, we will present a semi-automatic detection procedure for measuring NT thickness based on DP [8, 10, 17] improved by a nonlinear anisotropic diffusion filtering. We start by preprocessing an image, defining a region of interest in order to reduce interference from the image boundary and to adapt to different fetal head positions and sizes of NT layer. At this stage we also apply coherence-enhancing diffusion filtering, which uses an understanding at the structural features of NT images to enhance the NT region. This filtering improves the difficulty of border segmentation due to the speckle noises and weak edges in ultrasound images of fetus NT. In the main segmentation step, we use a DP procedure to determine the location of the NT border by minimizing a cost function, which is a weighted sum of terms which expresses local measurements of the echo intensity and the intensity gradient of the image, and a geometrical constraint on the shape of the NT borders.

The rest of this paper is organized as follows: in Sect. 2, we describe our image-acquisition techniques, the characteristics of NT images and the preprocessing stage that defines the region of interest and show how we enhance the image using a CED filter. We also describe the segmentation procedure that we use to detect the border of the NT layer, and our measurement procedure. The results from this method are evaluated in Sect. 3, and we draw some discussions in Sect. 4.

2 Materials and methods

Fetal nuchal translucency images were acquired using an Accuvix XQ (Medison Co. Ltd, South Korea) ultrasound scanner with a convex 2–6 and 4–7 MHz transducer. They were saved in DICOM format using the SonoView software, which bundled with the scanner.

2.1 Preprocessing

2.1.1 Image characteristics

Nuchal translucency refers to the normal subcutaneous fluid-filled space between the back of the neck of a fetus and the overlying skin [3]. Figure 1 shows a representative image of the NT and a schematic illustration of the echo zones (Z1–Z5) and the borders (B1, B2) of the NT layer. The NT thickness is defined as the maximum thickness of the translucent space (Z3) between the skin (Z4) and the soft tissue (Z2) overlying the cervical spine in the sagittal section of the fetus. Manual measurements of the NT layer are made by placing the crossbar of on-screen calipers on the inner edges of the two thin echogenic lines that border the NT layer. Measurement of the maximum thickness should be made with the calipers perpendicular to the long axis of the fetus. An ultrasonographic sagittal view of a fetus and the calipers is shown in Fig. 2a, and Fig. 2b illustrates correct and incorrect placement of the calipers on the anatomical structures of the fetus [12].
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig1_HTML.gif
Fig. 1

Echo zones and NT borders

https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig2_HTML.gif
Fig. 2

Using calipers to measure NT thickness. a Ultrasonographic sagittal view of a fetus with calipers; b placement of calipers on anatomical structures of the fetus

2.1.2 The region of interest

Nuchal translucency images can differ widely because of fetal movement during a scan of several minutes. Therefore, we need to define a region of interest (ROI) which reduces interference by the image boundary and which can compensate for changes in the fetal head position and the size of the NT layer. Since our ultimate purpose is to measure the thickness of the widest part of the NT layer, the user creates a rectangular region of interest that is sure to include the appropriate part of the layer.

Figure 3 shows a selection of NT images corresponding to different fetal positions. In Fig. 3a the fetal skin is clearly separate from the amnion, but in Fig. 3b there is no clear distinction between them. Fetal images with the head in upside-down or oblique positions are shown in Fig. 3c, d. By creating an appropriate ROI we can avoid interference from speckles, echo zones Z2 (soft tissue) and Z4 (skin).
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig3_HTML.jpg
Fig. 3

Nuchal translucency images of different fetal positions (the rectangle indicates the ROI). a A fetus image with a clear amniotic edge; b a fetus image with an indistinct amniotic edge; c a fetus in an upside-down position; and d a fetus in an oblique position

2.1.3 Coherence-enhancing diffusion

We now apply a CED filter, which emphasizes very thin structures within an NT image, even with a high level of noise, allowing us to detect the border accurately. CED filtering enhances the borders of the NT region while blurring the area inside the NT layer. CED filtering is based on the diffusion theory, which requires the solution of a partial differential equation (PDE). Locally coherent structures make the diffusion process more directional in both the gradient and the contour directions [1]. This is why CED filtering is appropriate for enhancing thin structures such as blood vessels or the fetus NT.

The CED filter uses a structure tensor (or structure matrix) to reduce speckle noise and enhance the image contrast [16], a technique, which derives from the more basic diffusion filter of Perona and Malik. The diffusion tensor used by the CED filter is obtained from the eigenvectors of the structure matrix, and the eigenvalues determine the strength of diffusion in the principal directions [1, 16]. The 2D structure matrix is calculated from the intensity gradient at each pixel, which represents the second-order derivatives. These can be expressed as follows:
$$ {\text{ }}{\left( {\begin{array}{*{20}c} {{I_{x} ^{2} }} & {{I_{x} I_{y} }} \\ {{I_{x} I_{y} }} & {{I_{y} ^{2} }} \\ \end{array} } \right)}. $$
(1)
The multiscale structure matrix is
$$ J(I) = {\left( {\begin{array}{*{20}c} {{j_{{xx}} }} & {{j_{{xy}} }} \\ {{j_{{xy}} }} & {{j_{{yy}} }} \\ \end{array} } \right)} = {\left( {\begin{array}{*{20}c} {{K_{\rho } \times I_{x} ^{2} }} & {{K_{\rho } \times (I_{x} I_{y} )}} \\ {{K_{\rho } \times (I_{x} I_{y} )}} & {{K_{\rho } \times I_{y} ^{2} }} \\ \end{array} } \right)}, $$
(2)
using a Gaussian convolution operator, which is defined as:
$$ K_{\rho } (x,y) = (2\pi \rho ^{2} )^{{ - 1}} \exp {\left( {\frac{{x^{2} + y^{2} }} {{2\rho ^{2} }}} \right)}. $$
(3)
These two equations can be reformulated using eigenvalue decomposition, where ω1, ω2 are eigenvectors and T signifies a transpose:
$$ J(I) = {\left( {\begin{array}{*{20}c} {{\omega _{1} }} & {{\omega _{2} }} \\ \end{array} } \right)}{\left( {\begin{array}{*{20}c} {{\mu _{1} }} & {0} \\ {0} & {{\mu _{2} }} \\ \end{array} } \right)}{\left( {\begin{array}{*{20}c} {{\omega _{1} ^{T} }} \\ {{\omega _{2} ^{T} }} \\ \end{array} } \right)}. $$
(4)
The first and second eigenvalues can then be calculated by solving the equation, which is created by setting the determinant of the structure matrix to zero:
$$ \begin{aligned}{} & x^{2} - (j_{{xx}} - j_{{yy}} )x - j_{{xy}} ^{2} = 0, \\ & \mu _{{1,2}} = \frac{1} {2}{\left( {j_{{xx}} + j_{{yy}} \pm {\sqrt {(j_{{xx}} - j_{{yy}} )^{2} + 4j_{{xy}} ^{2} } }} \right)}. \\ \end{aligned} $$
(5)
Local coherence is defined as the difference between the eigenvalues, and is calculated as follows:
$$ \mu _{1} - \mu _{2} = {\sqrt {(j_{{xx}} - j_{{yy}} )^{2} + 4j_{{xy}} ^{2} } }. $$
(6)
Coherence is greater in a region of high anisotropy, and such a region fulfils the irregularity condition of μ≥ μ>> 0, whereas μ≈ μ 0 is true elsewhere. The diffusion tensor and its decomposition using eigenvalues are defined in the following way:
$$ \begin{aligned}{} & \frac{{\partial I(x,y,t)}} {{\partial t}} = \;{\rm {div}}[D\nabla I] \\ & D(I) = {\left( {\begin{array}{*{20}c} {{\omega _{1} }} & {{\omega _{2} }} \\ \end{array} } \right)}{\left( {\begin{array}{*{20}c} {{\mu _{1} }} & {0} \\ {0} & {{\mu _{2} }} \\ \end{array} } \right)}{\left( {\begin{array}{*{20}c} {{\omega _{1} ^{\text{T}} }} \\ {{\omega _{2} ^{\text{T}} }} \\ \end{array} } \right)}, \\ \end{aligned} $$
(7)
where t is time and div is the divergence operator in the diffusion formula [13]. The elements of the tensor are
$$ \begin{aligned}{} & d_{{xx}} = \frac{1} {2}{\left( {c_{1} + c_{2} + \frac{{(c_{1} - c_{2} )(j_{{xx}} - j_{{yy}} )}} {\alpha }} \right)}, \\ & d_{{yy}} = \frac{1} {2}{\left( {c_{1} + c_{2} - \frac{{(c_{1} - c_{2} )(j_{{xx}} - j_{{yy}} )}} {\alpha }} \right)}, \\ & d_{{xy}} = {\left( {\frac{{(c_{2} - c_{1} )j_{{xy}} }} {\alpha }} \right)}, \\ & \alpha = {\sqrt {(j_{{xx}} - j_{{yy}} )^{2} + 4j_{{xy}} ^{2} } }. \\ \end{aligned} $$
(8)
The diffusion speeds c1, c2 depend on these eigenvalues and a constant k [16]:
$$ \begin{aligned}{} & c_{1} = \max (0.01,{\text{ }}1 - {\text {e}}^{{ - (\mu _{1} - \mu _{2} )^{2} /k^{2} }} ) \\ & c_{2} = 0.01. \\ \end{aligned} $$
(9)
Figure 4 shows the effects of a Gaussian filter, a median filter and a CED filter.
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig4_HTML.jpg
Fig. 4

Effects of various filters. a Original image; b Gaussian-filtered image; c median-filtered image; d CED-filtered image

2.2 Border detection

The accurate segmentation of the border is never going to be easy since almost all ultrasound images have a high level of speckle noise, other imaging artifacts and weak edges due to random scattering. We address this problem using a global optimization approach based on DP. We will now describe the segmentation method in detail.

2.2.1 Cost function

Cost functions are built for each of the borders of the NT layer, which we will call B1 and B2. If the ROI is an M × N rectangle, then all possible borders BN can be considered as polylines with N nodes:
$$ B_{{N\;}} \; = \;\{ {\user2 {p}}_{1} ,{\user2 {p}}_{2} , \ldots ,{\user2 {p}}_{{N - 1}} ,\;{\user2{p}}_{N} \} , $$
(10)
where the pixels pN–1 and pN are horizontal neighbors, and N is the horizontal length of a contour line. The cost function C(BN) is defined as a sum of local costs along a candidate border BN:
$$ C(B_{N} )\; = \;c_{f} ({\user2{p}}_{1} )\; + \;{\sum\limits_{i = 2}^N {(c_{f} ({\user2{p}}_{i} )\; + \;c_{g} ({\user2{p}}_{{i - 1\;}} ,\;{\user2{p}}_{i} ))} }. $$
(11)
At a point pi, the local cost is made up of two terms cf(pi) and cg(pi−1, pi), which are defined as follows:
$$ c_{f} ({\user2{p}}_{i} )\; = \;{\sum\limits_{j = 1}^k {w_{j} f_{j} ({\user2{p}}_{i} )} }\quad \quad \quad (i{\kern 1pt} \; = \;1, \ldots ,N), $$
(12)
$$ \begin{aligned}{} & c_{g} ({\user2{p}}_{{i - 1}} ,\;{\user2{p}}_{i} )\; = \;w_{{k + 1}} g({\user2{p}}_{{i - 1}} ,\;{\user2{p}}_{i} )\quad \quad \\ & g({\user2{p}}_{{i - 1\;}} ,\;{\user2{p}}_{i} ) = {\left| {d({\user2{p}}_{i} )\;{\kern 1pt} - \;d({\user2{p}}_{{i - 1}} )} \right|}^{2} \quad \quad (i{\kern 1pt} \; = \;2,\; \ldots,\;N), \\ \end{aligned} $$
(13)
where fj(pi) are image feature terms, k is the number of image features being considered, g(pi−1, pi) is a geometrical force term, d is vertical distance between the border being estimated and a reference line at a node and wj (j = 1, 2, 3, 4) is a weighting factor. We use three image features, so k is 3. The weighting factors are determined empirically for each border using the constraint |w1| + |w2| + |w3| + |w4| = 1.

The terms in the cost function associated with a particular type of border must reflect the characteristics of the image features in its vertical neighborhood and the geometrical form of the border. These cost terms are therefore defined such that a stronger image feature at pi will yield a lower local cost. The desired border corresponds to the BN which minimizes the cost function C(BN).

Figure 5 shows a 3D view of the NT layer. The border B1 (in the y-direction) is below a bright region, and above a dark region, while the border B2 is located above a bright region and below a dark region. That is, the characteristics of image feature of upper border B1 in vertical neighboring echo zones are opposite to those of lower border B2. Due to the differences of characteristics between both borders, therefore, weighting factors of image feature terms of lower border and those of upper border have the opposite signs, respectively. From these observations, we define cost terms with weighting factors that correspond to the characteristics of each border.
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig5_HTML.gif
Fig. 5

3D view of the nuchal translucency layer

We will now describe the cost terms used in the cost function for B2. The cost terms for B1 are similar to those for B2 except for the value of weighting factors.

The first term f1(pi) measures the average intensity of n (here n = 3) pixels below a pixel pi, with the aim of detecting a pixel that belongs to a line above a strong echo zone Z4. The second term f2(pi) measures the average intensity of m (here m = 2) pixels above a pixel pi, and favors a pixel that belongs to a line immediately below a dark NT space Z3. The third term f3(pi) measures the downward intensity gradient which is expected at the upper edge of an echo zone Z4. The intensity gradient is estimated as the vertical slope of intensity at the pixel pi using a vertical gradient operator [1 0 −1]T instead of a 5 × 5 neighborhood window used in [8, 10, 17]. The final cost term g(pi−1, pi) is proportional to the square of the difference in vertical distance between the border being estimated and a reference line at node pi; this term is designed to ensure border continuity. In the case of B2, g(pi−1, pi) is calculated with respect to a horizontal reference line. In estimating the cost term g(pi−1, pi) for B1, the reference line is the border B2 that has already been detected.

2.2.2 Dynamic programming

Dynamic programming, which is typically applied to optimization problems, can be used for border detection. The optimum border minimizes the cost function C(BN). The search for a solution must be performed globally in the region of interest so as to avoid local minima due to interference patterns such as speckles.

To recast our problem as a DP search, we begin by rewriting Eq. (11) in a recursive form as follows:
$$ \begin{aligned}{} & C(B_{1} )\; = \;c_{f} ({\user2{p}}_{1} )\;, \\ & C(B_{n} )\; = \;C(B_{{n - 1}} )\; + \;(c_{f} ({\user2{p}}_{n} )\; + \;c_{g} ({\user2{p}}_{{n - 1}},\;{\user2{p}}_{n} ))\quad \quad \quad (n\; = \;2,\ldots,N). \\ \end{aligned} $$
(14)
We denote the candidate minima of the cost function of polyline BN as \( \ifmmode\expandafter\tilde\else\expandafter\sim \fi{C}(B_{n} ). \) Applying Eq. (14), the multistage cost accumulation process can now be expressed as
$$ \begin{aligned}{} & \ifmmode\expandafter\tilde\else\expandafter\sim \fi{C}(B_{1} )\; = \;c_{f} ({\user2{p}}_{1} )\;, \\ & \ifmmode\expandafter\tilde\else\expandafter\sim \fi{C}(B_{n} )\; = \;{\mathop {\min }\limits_{{\user2{p}}_{{n - 1}} } }\;\{ \ifmmode\expandafter\tilde\else\expandafter\sim \fi{C}(B_{{n - 1}} )\; + \;(c_{f} ({\user2{p}}_{n} )\; + \;c_{g} ({\user2{p}}_{{n - 1}},\;{\user2{p}}_{n} ))\} \quad \;(n\; = \;2,\ldots,N). \\ \end{aligned} $$
(15)
To search for the optimum border, costs are accumulated following Eq. (15) and the location of pn−1, which defines \( \ifmmode\expandafter\tilde\else\expandafter\sim \fi{C}(B_{n} ), \) is stored in a pointer array at each stage n. The arrows in Fig. 6a point to the pixel in the previous column with the minimum accumulated cost, which allows the points on the corresponding border to be derived by back-tracking.
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig6_HTML.gif
Fig. 6

Border detection using dynamic programming a the backtracking process, and b detecting the border B2 in the NT image

We will now describe the cost accumulation and back-tracking process in detail within an M × N ROI. A vertical window of height M is used to scan the border from the start to the end-point at N horizontal positions. The locations of the start and the end-point are the upper-left and upper-right pixels of the ROI, respectively, and correspond to the left and right arrows in Fig. 6b. At each column n, a candidate minimum cost function is found for each point, and its cost is accumulated. By scanning all the columns, the lowest accumulated cost is located. The position of the minimum-cost point becomes the end-point of the border. Then back-tracking is performed, following the pointers, until the first column in the ROI is again reached. This creates our estimate of the border. B2 can be identified relatively easily since the nearer echogenic zone Z4 is stronger than Z2. Therefore we find B2 first. We then search for B1 using a smoothed version of B2 as the reference line, which forms a lower limit in the search for B1.

2.3 Measurements

We measure the distance between B1 and B2 using linear regression. The process consists of five steps. In the first step, we calculate midpoints between the upper and lower edges at the right and left neighbors of a given value of x. The second step is to find the line L1 that bisects the two edges by linear regression, and the third step is to find a line L2 that is orthogonal to this bisector. The fourth step is to fit lines L3, L4 to B1 and B2 respectively, again using linear regression for the same values of x. Finally, we calculate the intersection points between L2 and L3 and between L2 and L4, and then measure the distance between each pair of points. These distances are estimates of the minimum, average and maximum NT thickness. Figure 7 illustrates this method of measuring the NT thickness.
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig7_HTML.jpg
Fig. 7

Schematic illustration of the measurement method using linear regression

3 Results

The algorithm that we have described in this paper was implemented in Visual C++ on a Pentium IV with 1 GB of RAM.

We have applied our method to 640 × 480-sized fetus NT images obtained by transabdominal and transvaginal ultrasonography. Figure 8 compares established edge-detectors with our method. Using the Sobel or Canny edge-detectors’ results in discrete borders and the snake [18] is not correctly matched to the borders due to the high level of noise and the weak edges, whereas our method extracts continuous borders accurately. Figure 9 shows results from various fetus NT images using our method. Figure 10 shows the position of the maximum and minimum NT thickness displayed on the extracted borders as different-colored lines. The green and yellow lines represent the position of the maximum and minimum NT thickness, respectively.
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig8_HTML.jpg
Fig. 8

Comparison of various edge detectors a original image; b Sobel; c Canny; d edge-detection using dynamic programming; e initial snake contours; f edge detection using snake

https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig9a_HTML.jpg
https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig9b_HTML.jpg
Fig. 9

Experimental results. Left original image, right detecting the border of a fetal NT

https://static-content.springer.com/image/art%3A10.1007%2Fs11517-007-0225-7/MediaObjects/11517_2007_225_Fig10_HTML.jpg
Fig. 10

Display of the locations where maximum and minimum measurements were obtained (green line maximum NT thickness, yellow line minimum NT thickness)

For quantitative evaluation of the proposed method, the experiments were performed on 30 fetus NT images. We calculated their correlation between semi-automatic and manual measurements for the maximum NT thickness. The degree of correspondence between both methods is indicated by the correlation ca, m , which is defined as follows:
$$ c_{\text{a},\text{m}} \; = \frac{{\text{Cov}}_{\text{a},\text{m}}}{\sigma _{\text{a}}\sigma _{\text{m}}}, $$
(16)
where Cova,m is the covariance between the semi-automatic and manual measurements and σa, σm are the standard deviations of each. The resulting values of the parameters μa, μm, σa, σm, ca,m are shown in Table 1.
Table 1

Comparison between semi-automatic and manual methods

 

Semi-automatic measurement, μa ± σa (mm)

Manual measurement, μm ± σm (mm)

Correlation ca,m

NTmax

2.81 ± 0.79

2.79 ± 0.77

0.99

4 Discussion

We have proposed a semi-automatic method of measuring NT in fetal ultrasonic scans, which combines a preprocessing step based on a nonlinear anisotropic diffusion filtering and DP procedure. Our method overcomes the speckle noise and weak edges in ultrasound image by applying a CED filter, which enhances the borders of the NT while blurring the NT layer itself. The proposed segmentation method then detects the borders of the NT by constructing a cost function that includes weighted cost terms to represent multiple image features such as intensity, gradient and border continuity. In addition, we display the locations where measurements of maximum and minimum NT thickness were obtained.

In our study, we have obtained good results without requiring manual correction in most cases. In particular, our method has overcome the problem of border continuity, which occurs when weak edges are disrupted by the scattering inevitably present in ultrasound images.

Our semi-automated measurements were compared with manual measurements, and were found to be equally accurate. In any case an automated approach naturally reduces the problems of operator variability, and hence reproducibility, that are always present with manual measurements. However, the process of measuring NT thickness basically has two limitations. The first limitation is to select the appropriate scanning plane. It is inherent in two-dimensional measurements in ultrasonography since the imaging plane does not always coincide with the required plane, that is, the true mid-sagittal section of the fetus. Both automatic and manual measurements may thus produce underestimates or overestimates of the NT thickness. The second limitation relates to the method itself due to fetal orientations. The proposed method may be only applied when fetal position is as horizontal as possible. However, the limitation can be explained by one of the standards for measuring NT [2], which a mid-sagittal section of the fetus should be obtained with the fetus horizontal on the screen.

For overcoming the limitations, we will investigate techniques to select the optimum NT view from a 3D data set. In addition, for clinical purposes some interactive tools may still be needed in order to correct some residual detection errors in extremely poor images.

Acknowledgments

This work is financially supported by the Ministry of Education and Human Resources Development (MOE), the Ministry of Commerce, Industry and Energy (MOCIE) and the Ministry of Labor (MOLAB) through the fostering project of the Lab of Excellency, and by the Korea Institute of Science and Technology Evaluation and Planning (KISTEP), under the Real Time Molecular Imaging program. We would like to thank the Medison Co. Ltd. for providing ultrasound image datasets for our experiments.

Copyright information

© International Federation for Medical and Biological Engineering 2007