Physics in Medicine & Biology

Related content

PAPER

Estimation of prenatal aorta intima-media thickness from ultrasound examination To cite this article: E Veronese et al 2014 Phys. Med. Biol. 59 6355

- Magnetic resonance imaging of the blood flow response to oral glucose J J Totman, L Marciani, S Foley et al. - Scan--rescan intra-observer variability of carotid atherosclerosis Arvin Vidal, Yves Bureau, Trevor Wade et al. - Automated CAILRS Kristen M Meiburger, Filippo Molinari, U Rajendra Acharya et al.

View the article online for updates and enhancements.

Recent citations - Giacomo Tarroni et al - Status and potential clinical value of a transthoracic evaluation of the coronary arteries Fabien Labombarda et al

This content was downloaded from IP address 139.184.14.159 on 18/09/2017 at 14:20

Institute of Physics and Engineering in Medicine Phys. Med. Biol. 59 (2014) 6355–6371

Physics in Medicine and Biology doi:10.1088/0031-9155/59/21/6355

Estimation of prenatal aorta intima-media thickness from ultrasound examination1 E Veronese2, G Tarroni2, S Visentin3, E Cosmi3, M G Linguraru4 and E Grisan2 2

  Department of Information Engineering, University of Padova, Via Gradenigo 6/b,35100 Padova, Italy 3   Department of Woman and Child Health, University Hospital of Padova, Via Giustiniani 3, 35128 Padova, Italy 4   Children’s National Medical Center Sheikh Zayed Institute for Pediatric Surgical Innovation, 111 Michigan Avenue NW, Washington, DC 20010, USA E-mail: [email protected] Received 19 November 2013, revised 21 April 2014 Accepted for publication 2 September 2014 Published 8 October 2014 Abstract

Prenatal events such as intrauterine growth restriction and increased cardiovascular risk in later life have been shown to be associated with an increased intima-media thickness (aIMT) of the abdominal aorta in the fetus. In order to assess and manage atherosclerosis and cardiovascular disease risk in adults and children, in recent years the measurement of abdominal and carotid artery thickness has gained a growing appeal. Nevertheless, no computer aided method has been proposed for the analysis of prenatal vessels from ultrasound data, yet. To date, these measurements are being performed manually on ultrasound fetal images by skilled practitioners. The aim of the presented study is to introduce an automatic algorithm that identifies abdominal aorta and estimates its diameter and aIMT from routine third trimester ultrasonographic fetal data. The algorithm locates the aorta, then segments it and, by modeling the arterial wall longitudinal sections  by means of a gaussian mixture, derives a set of measures of the aorta diameter (aDiam) and of the intima-media thickness (aIMT). After estimating the cardiac cycle, the mean diameter and the aIMT at the end-diastole phase are computed. Considering the aIMT value for each subject, the correlation between automatic and manual end-diastolic aIMT measurements is 0.91 in a range of values 0.44-1.10  mm, corresponding to both normal and pathological 1

Preliminary results have been presented in E Veronese, E Cosmi, S Visentin, E Grisan: 'Semiautomatic estimation of fetal aorta intima-media thickness from ultrasound examination', MICCAI Workshop on Perinatal and Paediatric Imaging: PaPI 2012. 0031-9155/14/216355+17$33.00  © 2014 Institute of Physics and Engineering in Medicine  Printed in the UK & the USA

6355

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

conditions. The automatic system yields a mean relative error of 19%, that is similar to the intra-observer variability (14%) and much lower that the interobserver variability (42%). The correlation between manual and automatic measurements and the small error confirm the ability of the proposed system to reliably estimate aIMT values in prenatal ultrasound sequences, reducing measurement variability and suggesting that it can be used for an automatic assessment of aIMT. Keywords: fetal ultrasound, image segmentation, mixture models, B-mode ultrasound, IUGR (Some figures may appear in colour only in the online journal) 1. Introduction According to the Barker hypothesis (Barker 1997) several adult diseases originate through adaptations of the fetus when it is undernourished. Low birth weight, caused either by preterm birth and/or intrauterine growth restriction (IUGR), was recently shown to be associated with increased rates of cardiovascular disease and non-insulin dependent diabetes in adult life (McMillen et al 2005, Hemachandra et al 2007, Langley-Evans et al 2009). It is well established that infants who had IUGR have a thicker aorta, suggesting that prenatal events (e.g. impaired fetal growth) might be associated with structural changes in the main vessels (Skilton et al 2005). Hence, the measurement of abdominal aortic intima-media thickness (aIMT) in children becomes a sensitive marker of atherosclerosis risk (Skilton et al 2005, Litwin and Niemirska 2009). Recently, a follow-up study showed that aIMT measurements in IUGR fetuses were inversely related to estimated fetal weight, showing that low birth weight and Doppler abnormalities may be correlated to an altered vascular structure causing possible endothelial damage (Cosmi et al 2009, Visentin et al 2013), consistent with the finding that atherosclerosis begins to develop first in the intima of the aorta (McGill et al 2000). On this basis, measurements of the abdominal aIMT from ultrasound (US) images have the potential to become powerful instruments for the assessment of fetal risk of atherosclerosis. Abdominal aIMT was defined as the distance between the leading edge of the blood-intima interface and the leading edge of the media-adventitia interface on the far wall of the vessel (Koklu et al 2007) (see figure 1 for a hystological section and figure 2 for a representative fetal abdominal aorta US frame). The increasing need for the assessment and management of atherosclerosis and cardiovascular disease risk resulted in a growing interest in the intima-media thickness as a biomarker for predicting metabolic disorders and cardiovascular risk in children and adults. Despite the fact that the alterations in the aorta are an early and sensitive marker of atherosclerotic risk (Jarvisalo et al 2001, Dawson et al 2009, Dulac et al 2011), the main attention has been devoted to the analysis and quantification of alterations of the carotid artery (CA) both in children (Dawson et al 2009, Litwin and Niemirska 2009, Jouret et al 2011, Crispi et al 2012, Dratva et al 2013, Thompson et al 2013) and in adults (Simon et al 2002, Touboul et al 2004, van der Meer et al 2004, Gaitini and Soudack 2005, De Groot et al 2008), since it allows an easy and high-resolution imaging. The widespread measurement of CA-IMT in adult patients has raised the need for quantitative and automatic measurement of CA-IMT (for a review see Molinari et al (2010)). On the contrary, the automatic measurement of aIMT in prenatal US images has been rarely attempted. Even if the problems of measuring CA-IMT or aIMT in adults and children and measuring aIMT in the fetus are similarly posed, the latter adds 6356

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 1.  Hematoxylin-eosin histological image of a normal abdominal aorta of a fetus at 28 weeks of gestation, evidencing the tunica intima (1), media (2) and adventitia (3) and the intima-media thickness (aIMT).

several challenges to the difficulties and limitations of measurements in the adult carotid and in the echogenic differences characteristics of the carotid and the aorta. In CA US examination of adults and children, the vessel position is almost fixed and it occupies the vast majority of the image. Moreover, CA is aligned with the image plane, making the estimation of the lumen position relatively robust. Finally, given the superficial position of the CA relatively to the body surface, the resolution obtained imaging the CA and its wall structures (intima, media and adventitia) is large compared to noise. In fetal US, on the contrary, the image resolution is lower than in adult CA examinations, since the structures to be imaged are deep within the maternal womb; this limitation is coupled with the reduced dimension of the vessels and of the intima and media layers of the fetus. Secondly, fetal movements cause the position and orientation of the abdominal aorta to be largely unpredictable: the aorta can move in and out of the US imaging window during the examination (see figure 3). Thirdly, the identification of the aorta is further complicated by the presence of the surrounding abdominal tissues and organs, since they typically appear at US examination as echogenic tissues enclosing hypoechogenic media (e.g. the appearance of the gastro-intestinal tract or bladder). To date, aIMT measurement has been performed manually by skilled practitioners, thus being susceptible to intra- and inter- operator variability. Moreover, in order to reduce the measurement variability and to produce results comparable for longitudinal and multicentric studies, the aIMT has to be measured in a consistent manner at the same point of the cardiac cycle. Thus, all measurements are manually taken at end-diastole which was identified as the time of maximal expansion of the vessel during the entire cardiac cycle (Litwin and Niemirska 2009). To the best of our knowledge, the only automatic method proposed in literature has been proposed in Veronese et al (2012), where the aorta was segmented on a small number of videos with an active-contour approach. Similarly to what has been proposed in literature for adult vessel analysis, the gradient information across the aorta boundaries is analyzed to estimate intima-media thickness. The reported correlation between automatic and manual measurement was 0.9 and the mean absolute error was of 0.95 pixels with respect to the manual measurement. Unfortunately the variability of aorta lumen and 6357

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 2.  Example of three US images showing manual measurement of aortic diam-

eter (dotted line) and aIMT (small continuous lines) in a fetal echography. Different acquisitions of various quality are shown.

Figure 3.  Example of a sequence of images from a cine-loop abdominal aorta acquisition at times t1 = 0.83 s (leftmost panel), t2 = 1.23 s (central panel) and t3 = 1.62 s (rightmost panel), where the aorta disappearing from the imaged echographic window is due to movement of the fetus or of the operator.

vessel boundaries appearance pose an overwhelming challenge to the settings of the active contour parameters. To address the challenges associated with manual measurements of small fetal vessels from US images we propose an automatic algorithm that identifies the abdominal aorta and estimates its diameter. Our computer-aided image analysis tool automatically estimates the cardiac cycle to provide consistent measurements of aIMT. This technique is intended to make the measurement of aIMT faster, accurate and reproducible. 2. Materials Thirtyfive video sequences were obtained from women undergoing routine US examinations during pregnancy, at the Department of Woman and Child Health of the University Hospital of Padova (Padova, Italy). The study was approved by the local ethical committee and all patients gave written informed consent. Fetal US data was acquired at a mean gestational age of 32 weeks (range 30 to 34 weeks) using a US machine equipped with a 5 to 7.5 MHz linear array transducer (Antares, Siemens Medical Solutions, Mountain View, CA), with a 70° FOV, image dimension 640 × 480 pixels and a variable resolution between 0.05 and 0.1 mm/pixel. Twentyfive fetuses were with appropriate weight for gestational age (AGA), 7 were AGA but 6358

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

with a gestational diabetic mother, 1 was small for gestational age (SGA), and 2 were affected by intrauterine growth restriction (IUGR) as previously defined (Visentin et al 2013). The localization of the abdominal aorta was performed in a sagittal view of the fetus at the dorsal arterial wall of the most distal 15 mm of the abdominal aorta sampled below the renal arteries and above the iliac arteries (Skilton et al 2005, Cosmi et al 2009). Gain settings were used to optimize image quality. After the localization, in order to measure the intima media thickness and diameter, the vessel was visualized in a longitudinal view of the fetus, making sure it was a maximal longitudinal section of the vessel (it contains the vessel diameter) and tilting the transducer to obtain an angle of insonation as close to 0° as possible and always less than 30°. Aortic diameter and aIMT measurements were manually determined at the time of the examination (online) by a skilled practitioner (either E.C. or S.V.), who subsequently reviewed the acquired videos at a workstation (offline procedure), blind to the results of the automatic system and to the clinical information on the fetuses. For the offline review, all frames corresponding to the end-diastolic phase were collected from the 35 videos (a total of 240 different frames) and those that were judged of sufficient quality with respect to the visibility of aortic wall were manually annotated to provide the ground truth. Three videos did not present any frame judged of sufficient quality (for a total of 22 frames) and were discarded during the off-line review. For the remaining 32 videos the total annotated number of frames was 165, resulting in a mean number of annotated end-diastolic frames per video of 5.15 and a mean number of frames judged of insufficient quality by the reader of 1.64 per video. The first 20 available frames were used to set up the algorithm parameters as will be described in the Parameters setting section; the parameters were then kept fixed. 3. Methods In order to correctly identify aIMT, we designed an algorithm that, starting from the acquired US video, firstly identifies an approximate skeleton of the aorta and then uses its central position as starting point to segment the vessel in each frame and finally estimate its diameter. After estimating the mean diameter of the aorta in each frame, heart rate and cardiac cycle are derived through the analysis of the diameter variation with time and the frames corresponding to the end-diastole cardiac phase are selected. On these frames, the regions along the vessel walls providing enough contrast for estimating the aIMT are identified and there the aIMT are measured. The diagram of the proposed processing pipeline is represented in figure 4 3.1. Initialization

In order to identify a reliable starting point for the vessel segmentation procedure and to provide an approximate estimate of the aorta longitudinal extension, we exploit the fact that each video was recorded so that the vessel was visualized in a longitudinal view of the fetus. This is necessary to identify the correct region in videos with a poor SNR or where other vessels or elongated hypoechogenic structures are present (see for example figure 1 central and right panel). In this way the aorta appears as a dark horizontal or slightly oblique area. We design a base filter K(x, y) of dimension L × L(L = 50 pixels in our specific implementation) to match the appearance of an aorta horizontally oriented in the image, obtained combining a dark large elongated area corresponding to the aorta lumen K1 and the bright boundaries corresponding to the vessel walls K2 and K3:  1 (x, y ) = e−0.5( σ ) K y

2

(1a) 6359

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 4.  Processing pipeline for a video of a fetal aorta. After the initialization and the tracking of the aortic centerline and diameter in each frame, the cardiac cycle is estimated and end-diastolic frames are selected; on these ones the region for aIMT estimation is identified, and a set of measurements is computed for each frame. The mean value of the aIMT measurements is provided as final value. ⎛ y + 1.5σ ⎞2 ⎟ 0.25σ ⎠

(1b)

⎛ y − 1.5σ ⎞2 ⎟ 0.25σ ⎠

(1c)

 (x, y ) = e−0.5⎜⎝ K 2

 (x, y ) = e−0.5⎜⎝ K 3

( Lx/3 )

 (x, y ) = (−K1 (x, y ) + K2 (x, y ) + K3 (x, y )) e−0.5 K

2

(1d)

For each video, we filter each frame with a bank of filters at multiple scales and multiple orientations, obtained from K(x, y) varying the scaling factor σ in the interval [5, 20] pixels and rotating with angles from −2π/15 to 2π/15 (figure 5). For each pixel, the maximum output of the filter banks is computed and the resulting image Imax(x, y) is binarized by means of a statistical threshold Imax(x, y) > θ,with θ = μ Imax + σImax , where μ Imax and σImax are the image mean and standard deviation, respectively. The binarized image is then eroded and then each connected component receives a score equal to its 6360

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 5.  Example of the directional filters used, shown with different orientation and different scales. Top row corresponds to scale parameter σ = 10, bottom row to scale parameter σ = 15, from left to right a rotation of the filter of 24°, 16°, 8° and 0°.

eccentricity, assumed to represent its vesselness. The scores from each frame are summed and the connected component having the highest accumulated vesselness score is taken as the approximated aortic region. From this region, the skeleton is estimated so to provide a rough aorta centerline, whose center (xc, yc) and direction θc are used as seed for the segmentation of the aorta. 3.2.  Aorta segmentation

In each frame, the central point of the skeleton is considered as a reliable seed point, i.e. the pixel from which the tracking of the vessel starts. The tracking method employed here is a customization of the one previously proposed (Grisan et al 2004): each point pi is described by the quaternion pi = (xi, yi, di, θi), where xi and yi are the point coordinates, di is the diameter of the vessel evaluated in (xi, yi) and θi is the direction, perpendicular to di. Starting from pi, the subsequent point pi + 1 is found by coupling both a priori and a posteriori information. The prior is represented by the fact that the direction θi + 1, the diameter di + 1 and the gray level intensity of adjacent points have to be similar, or at least smoothly varying. The a priori information provides a predicted vessel point pi +̂ 1, considering that the versor v = (cos (θ), sin (θ)) represents the most likely direction for a new vessel point pi +̂ 1 to be found, starting from pi. Given a constant tracking step s,

(xi+̂ 1, yi+̂ 1) = (xi, yi ) + s · v

(2)

the information on this estimated new point, on its diameter and direction, are then refined with the a posteriori information obtained employing a fuzzy c-mean clustering (Bezdek 1981) procedure on the intensity profile perpendicular to the vessel direction. The clustering procedure assigns to each pixel on the line centered in pi +̂ 1 and with direction θi + 1 + π/2, a probability of belonging to the ‘dark’ (corresponding to the vessel lumen) and ‘bright’ (corresponding to the soft tissues) classes, depending on its intensity value. By thresholding the estimated probability (we used a value of 0.7 (Grisan et al 2004)), the pixels on the line are 6361

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 6.  Centerline of the aortic lumen superimposed to the analyzed frame (dotted line), together with the estimated aorta diameters at points along the centerline (solid lines).

segmented into vessel lumen or background, so that the center of the vessel-related cluster becomes the refined estimate pi + 1 and the refined direction becomes θi + 1 = pi + 1 − pi. At variance with the original method, where a small intensity difference between the identified cluster centers was used as stopping criterion, here we allow the update of the center pi +̂ 1 and diameter di +̂ 1 even in presence of uninformative posterior, setting pi + 1 = pi +̂ 1, until the data-driven posterior is recovered. This procedure allows on the one hand to cope with the variability of US data and on the other with the effects of fetal movement that may results in the vessel lumen moving out of the imaging window for a small number of frames, or in a portion of the image (see e.g. figure 3). 3.3.  Cardiac phase estimation

Since measurement of intima-media thickness has to be obtained only at the end-diastolic phase of the cardiac cycle, we need to correctly identify the correspondent frames from the information on the aorta diameter, whose dimension varies in phase with the heart beat as it can be seen in figure 7. For each analyzed frame k the tracking procedure provides the aorta centerline pk̂ and diameter dk̂ (figure 6), so that the mean diameter value μaD (k ) = E [dk̂ ] for each frame can be computed. The variation of μaD(k) in time is modeled as a sinusoidal function aDiam (t) = A· sin (ft + ϕ), that is fit on the data μaD(k) by solving a weighted non linear least-squares problem (figure 7), obtaining the optimal value A ,̂ f ̂ and ϕ .̂ The times corresponding to the maxima of the estimated sinusoid model represent the end-diastolic cardiac phase corresponding to the frames k diastole = (π / 2 − ϕ )̂ / f ̂ + 2mπ / f ,̂ with m ∈ ℕ. 3.4.  Measurement region identification

In order to provide only reliable measurements of the intima-media thickness, the regions along the aorta where the vessel wall and its structure are most visible are automatically identified in each diastolic frame. For each point along the vessel wall, the contrast is computed as the difference between the mean gray level within a region 0.25d ̂ wide on the interior of the vessel border and the mean gray level within a region 0.25d ̂ wide on the exterior of the vessel border. The set of points where the contrast is larger than θMR of the maximum intensity value in the frame (figure 8) compose the measurement region MR. 6362

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Aorta diameter [mm]

2.6

2.4

2.2

2

1.8

1.6 0

0.5

1

1.5

2

2.5

time [s] Figure 7. Mean aorta diameter (solid line) and estimated sinusoidal fit aDiam(t)

(dashed line), from a representative US video.

If a frame results with an empty MR, the corresponding aIMT value is considered as a missing value, instead of being estimated through the procedure described in the following sections.

3.5.  Abdominal aIMT estimation

To detect the aIMT, for each point pi in MR we use the gray level profile centered in (xi, yi), π with direction ± −∣θi∣ and length equal to (1.5 · di)/2 pixels on both sides. We consider as 2 aIMT the distance between the leading edge of the blood-intima interface and the leading edge of the media-adventitia interface on the far wall of the vessel (Koklu et al 2007). Since the edges we are interested in are rarely well defined, but rather appear as a local intensity minimum between two bright peaks related to the intima-media and adventitia layers, we choose to model the profile intensities comprising lumen, intima-media and adventitia as a mixture of three Gaussians, at variance with the more complex model proposed for adult carotid that uses a mixture of three Nakagami distributions (Destrempes et al 2009). Our model becomes: 3

GM  (g) = ∑ zj · e

⎛ g − μj ⎞2 −0.5· ⎜ σ ⎟ ⎝ j ⎠

(3)

j=1

where μj is the mean of the jth Gaussian, σj its standard deviation and zj its mixing proportion. The mixture is usually composed of a large-variance gaussian G1(g), possibly accomodating both the background signal within the lumen and the low-frequency variation of the intensities across the aorta edge and two small-variance components G2(g) and G3(g), modeling the vessel wall layers. The parameters are estimated by fitting GM (g) to the gray level profile, using a gradient-based nonlinear least squared minimization. By this means, we can identify the 6363

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 8. Automatically identified region of the aorta for a reliable estimate of the

aIMT (starred points on the skeleton and framed region).

0.7

0.7

aIMT

0.6 0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0

aIMT

0.6

0

20

40

60

80

100

120

140

160

180

200

0 0

20

40

60

80

100

120

140

160

180

200

Figure 9. Mixture of Gaussian fit on the gray levels across the aorta wall in two representative cross-sections of different images, with a successful estimate of the aIMT.

media-adventitia (MA) interface as the local minimum gMA ̂ of GM (g) in the interval g ∈ [μ2, μ3]. In order to obtain a reliable measurement, we further impose that: ̂ ) , GM (μ3) − GM (g )̂ MA ⩾ Trel· max gGM (g) min GM (μ2 ) − GM (gMA

(

)

(4)

where Trel is a user defined threshold used to set the sensitivity to small variations in the intensity profile as fraction of the maximum intensity of the profile (see figures 9 and 10). We then refine the blood-intima (BI) interface position by evaluating the point gBÎ for which GM (gBÎ ) = 0.5 * GM (μ1), in the interval g ∈ [0, μ1]. For the kth frame and for each of the pi points, i = 1, …, P in the ROI, we thus obtain a value for the aIMT computed automatically: aIMT ̂ , k, p − gBÎ , k, p auto (k , p ) = gMA 

(5)

From the set of P measurements aIMTauto(k, p) in the frame k, we can compute their sample mean to obtain a single value aIMTauto(k) for the frame under analysis. 6364

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0

0

20

40

60

80

100

120

140

160

180

200

0

0

20

40

60

80

100

120

140

160

180

200

Figure 10.  Mixture of Gaussian fit on the gray levels across the aorta wall in two repre-

sentative cross-sections of different images, where the estimated tonaca media boundary were rejected by the application of the condition described in equation (4). 0.28

0.1

0.26

0.2

0.24

0.3

0.22 rel

0.4

T

0.2 0.5 0.18 0.6 0.16 0.7

0.8 0.1

0.14

0.2

0.3

0.4

0.5

θ

0.6

0.7

0.8

0.9

MR

Figure 11.  RMSE between the automatic and manual aIMT measurements for different

combinations of the parameters θMR and Trel.

3.6.  Parameters setting

Throughout the analysis of the fetal aorta videos, there are two parameters that can vary the sensitivity of the proposed procedure. The first, θMR introduced for the definition of the measurement region MR, is used to increase, in each frame, the number of accepted measurement sites along the vessel boundaries, but possibly introducing unreliable measurement. The second, Trel, has been introduced when estimating the aIMT and is used to vary the sensitivity to the presence or absence of the intensity valley between the intima-media and the adventitia. A grid search on the two parameters has been performed, testing each possible couple of parameter values in the range [0, 0.1, ···, 0.9, 1] on the first 20 frames of the annotated data set. For each couple, the root mean squared error (RMSE) between the automatic aIMT and the manual aIMT measurement is computed. The combination of parameters yielding the smaller root mean squared error was selected and used for the analysis of the remaining data. The value of the θMR has been set to 0.5, whereas the value of Trel has been set to 0.3 (figure 11). 6365

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Table 1.  Manual aIMT variability for the different frames in the dataset comparing:

errors and correlation among values measured by the same operator at different times (intra-observer variability), errors and correlation among values measured by the two operators (inter-observer variability). Single frame aIMT manual variability

Intra-observer variability Inter-observer variability

Mean error [mm]

Mean error %

RMSE [mm]

Correlation

0.10 0.31

12% 41%

0.16 0.35

0.56 0.51

Table 2.  Manual aIMT variability for the different patients in the dataset comparing:

errors and correlation among values measured by the same operator at different times (intra-observer variability), errors and correlation among values measured by the two operators (inter-observer variability) and errors and correlation among values measured online during the examination procedure using the tools provided by the echographic device and those measured offline reviewing the video data. Subject aIMT manual variability

Intra-observer variability Inter-observer variability Online-offline variability

Mean error [mm]

Mean error %

RMSE [mm]

Correlation

0.10 0.30 −0.16

14% 42% 55%

0.12 0.32 0.26

0.78 0.76 0.06

Table 3.  Errors and correlation comparing the automatic aIMT results with respect the manual measurements from operator 1. The first row reports the statistics for the measurements obtained in each frame in the data set, whereas the second row reports the error statistics comparing mean automatic and mean manual measurements for each subject.

Automatic results

Frame aIMT Subject aIMT

Mean error [mm]

Mean error % RMSE [mm]

Correlation

0.15 0.13

24% 19%

0.68 0.91

0.05 0.02

4. Results 4.1.  Manual measurement variability

During the offline measurement procedure, we were able to collect for each frame k in the data set a manual aIMT measurement from operator 1 (aIMTop1(k)), a manual aIMT measurement from operator 2 (aIMTop2(k)) and a second manual measurement from operator 1 (aIMTop1,rep(k)) performed 3 months after the first, blind to the first manual annotation. We can then evaluate the measurement variability within and between operators comparing the aIMT values in term of mean error, mean percentage error, root mean squared error (RMSE) and correlation between the operators. The results are reported in table 1. Moreover, since the aim of the procedure is to provide a reliable aIMT measurement for a single patient s, we can lump together the manual annotations belonging to each specific 6366

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Difference between manual and automatic aIMT [mm]

0.3 0.2 0.1 0 −0.1 −0.2 −0.3 −0.4 −0.5 −0.6

Difference between manual and automatic subject aIMT [mm]

0.2

0.4

0.6 0.8 1 1.2 Average between manual and automatic aIMT [mm]

1.4

1.6

0.3 0.25 0.2 0.15 0.1 0.05 0 −0.05 0.4

0.5

0.6 0.7 0.8 0.9 1 1.1 Average between manual and automatic subject aIMT [mm]

1.2

1.3

Figure 12.  Bland-Altman plots showing the agreement between the set of manual and automatic aIMT measurements. The top panel shows the plot for all 160 frames in the manually annotated dataset: with the exception of only 6 frames, all difference values lie within the ± 1.96σ (dotted lines) range. The bottom panel shows the plot for all 35 subjects in the manually annotated dataset: all differences lie within the ± 1.96σ (dotted lines) range.

patient aIMT (k, s), so to obtain the subject intima-media thickness aIMT (s) computing the average value of all lumped aIMT (k, s). Thus, we have the measurements from the two operator aIMTop1(s) and aIMTop2(s), the repeated measurement aIMTop1,rep(s) and the additional value aIMTonline(s) representing the  measurement evaluated during the data acquisition, i.e. while visiting the patient, with the measurement tool provided by the echographic device. We report the measurement variability at the patient level between- and within-operators and the variability between the measurements obtained offline, i.e. when reviewing the data and online, i.e. while examining the patient. Results are reported in table 2.

6367

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Automatic aIMT per subject [mm]

1.2 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

Manual aIMT per subject [mm]

Figure 13.  Manual versus automatic estimation of aIMT on the 35 patients in the data set. The Pearson correlation between the manual and automatic measurements is 0.91 and the regression line is reported in the plot (dotted line).

4.2.  Automatic measurement accuracy

Following the same analysis performed for the manual aIMT variability and taking the manual measurements aIMTop1 as ground truth, we compare the results from the automatic procedure both at the frame level aIMTauto(k) and at the subject level averaging the measurements aIMTauto(k, s) belonging to the same subject s. Summary statistics of the errors and the correlation between automatic and manual measurements are reported in table 3. The statistics show an agreement between the ground-truth and the automatic aIMT that is equivalent to the intra-observer variability, both when comparing the results at the frame level and at the subject level. This is further confirmed by the Bland-Altman plots reported in figure 12. Additionally, the correlation analysis for the subject-level aIMT measurements is reported in figure 13 and all values of the manual aIMTop1(s) and automatic aIMTauto(s) are reported in table 4. Finally, representative results on end-diastolic frames from videos corresponding to the manual annotation shown in figure 1 are reported in figure 14. 5. Discussion The present study developed an automated system to measure fetal aIMT in order to reduce inter- and intra- observer variability. Despite the growing attention to abdominal aorta thickness as an early and sensitive marker of atherosclerosis in children and adults alike, no study has been devoted to the automatic analysis of aIMT on clinical ultrasound data yet. Less more so in fetal echography, even if aIMT has been linked to a higher risk of metabolic syndrome in later life. The proposed automatic measurement has the privilege to be potentially performed by all operators, with only two parameters that might be tuned, namely the contrast threshold to define the measurement region and the value Trel to discard unreliable minima in the lumenintima-media-adventitia gray-level profile. From a clinical perspective, it potentially allows scheduling which fetus should be followed-up after delivery during third trimester routine ultrasound. Moreover, previous studies 6368

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Table 4.  aIMT values obtained during the examination (online), those reviewed by operator 1 and used as ground-truth and the automatic estimate.

aIMT subject measurements Subject

Online

Operator 1

Automatic

Subj. 1 Subj. 2 Subj. 3 Subj. 4 Subj. 5 Subj. 6 Subj. 7 Subj. 8 Subj. 9 Subj. 10 Subj. 11 Subj. 12 Subj. 13 Subj. 14 Subj. 15 Subj. 16 Subj. 17 Subj. 18 Subj. 19 Subj. 20 Subj. 21 Subj. 22 Subj. 23 Subj. 24 Subj. 25 Subj. 26 Subj. 27 Subj. 28 Subj. 29 Subj. 30 Subj. 31 Subj. 32 Subj. 33 Subj. 34 Subj. 35

0.700 0.900 0.450 0.700 0.650 0.650 0.500 0.500 0.500 0.500 0.300 0.250 0.550 0.350 0.350 0.350 0.350 0.450 0.450 0.450 0.450 0.600 0.450 0.600 0.500 0.500 0.850 0.450 0.250 0.250 0.550 0.450 0.450 0.500 0.500

0.752 0.827 0.707 — 0.748 0.782 0.676 0.588 1.082 0.839 0.643 0.802 0.586 0.811 0.741 1.109 0.673 0.539 0.591 0.655 0.709 — 0.510 0.624 0.645 0.649 — 0.439 0.615 0.424 0.615 1.032 0.689 0.518 0.461

0.954 0.937 0.970 0.990 0.828 0.895 0.839 0.830 1.153 1.080 0.693 0.952 0.746 1.004 0.958 1.234 0.904 0.625 0.659 0.878 0.706 0.877 0.696 0.643 0.877 0.787 1.008 0.489 0.633 0.442 0.864 1.268 0.760 0.589 0.652

have shown that in IUGR infants, aIMT was greater in those with the lowest birth weight. Such evidence suggests that atherogenesis and an increased arterial stiffness may be a potential mechanism mediating the mentioned epidemiological link between impaired fetal growth and later cardiovascular disease in adulthood, similar to the major environmental risk factors, such as cigarette smoking or hypertension. Unlike these studies, which focus on older age groups, aIMT measurement during pregnancy allows for prospective examination of the effects of impaired fetal growth, without the effect of postnatal confounders, directly evaluating this known marker of endothelial dysfunction and potential arterial health directly in utero from the second-third trimester of gestation and in newborns (Cosmi et al 2009, Zanardo et al 2011). As a final note, abdominal aorta seems to be the one of the first vessels involved by atherosclerotic changes (Brenner et al 2006), so that aIMT is a sensitive marker of early disfunction, but multicentric studies and trials face the challenge of standardizing the aIMT measurement, that requires skilled and experienced operators to reduce variability in ultrasound acquisition and in image interpretation. In conclusion, the introduction of the proposed automatic system might have the advantage to acknowledge robustly and across multiple centres which fetuses 6369

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Figure 14.  Representative results of the automatic analysis on end-diastolic frames

extracted from the videos of the patients corresponding to the manual measurement shown in figure 1: aorta centerline (white dotted line), mean estimated diameter (white solid line with circle heads) and estimated intima-media at points in MR (white solid line with asterisk heads) are reported.

share the risk of cardiovascular dysfunction further in life and to provide repeatable quantitative imaging markers of endothelial dysfunction. 6. Conclusions The high values of correlation between manual and automatic results suggest that the proposed algorithm provides a reliable technique for a quick measurement of important structures during ultrasonographic fetal biometry, such as aorta diameter and aIMT. Besides, being fully automatic, it allows avoiding the problems of intra- and inter- operator variability, typical of any manually performed measurement. References Barker D J P 1997 Maternal nutrition, fetal nutrition and disease in later life Nutritions 13 807 Bezdek J C 1981 Pattern Recognition with Fuzzy Objective Function Algorithms (Dordrecht: Kluwer) Brenner D et al 2006 Cytokine polymorphisms associated with carotid intima-media thickness in stroke patients Stroke 37 1691–6 Cosmi  E, Visentin  S, Fanelli  T, Mautone  A J and Zanardo  V 2009 Aortic intima media thickness in fetuses and children with intrauterine growth restriction Obstet. Gynecol. 114 1109–14 Crispi  F, Figueras  F, Cruz-Lemini  M, Bartrons  J, Bijnens  B and Gratacos  E 2012 Cardiovascular programming in children born small for gestational age and relationship with prenatal signs of severity Am. J. Obstet. Gynecol. 207 e1–9 Dawson J D, Sonka M, Blecha M B, Lin W and Davis P H 2009 Risk factors associated with aortic and carotid intima-media thickness in adolescents and young adults: the muscatine offspring study J. Am. College Cardiol. 53 2273–9 de Groot E, van Leuven S I, Duivenvoorden R, Meuwese M C, Akdim F, Bots M L and Kastelein J J 2008 Measurement of carotid intima–media thickness to assess progression and regression of atherosclerosis Nat. Clin. Pract. Cardiovasc. Med. 5 280–8 Destrempes F, Meunier J, Giroux M, Soulez G and Cloutier G 2009 Segmentation in ultrasonic b-mode images of healthy carotid arteries using mixtures of nakagami distributions and stochastic optimization IEEE Trans. Med. Imaging 28 215–29 Dratva J, Breton C V, Hodis H N, Mack W J, Salam M T, Zemp E, Gilliland F, Kuenzli N and Avol E 2013 Birth weight and carotid artery intima-media thickness J. Pediatr. 162 906–11 Dulac Y, Tauber M and Jouret B 2011 Aortic or carotid intima-media thickness to evaluate children born small for gestational age? Horm. Res. Paediatr. 77 340 6370

E Veronese et al

Phys. Med. Biol. 59 (2014) 6355

Veronese E, Poletti E, Cosmi E and Grisan E 2012 Estimation of prenatal aorta intima-media thickness in ultrasound examination Proc. SPIE Med. Imag. Gaitini  D and Soudack  M 2005 Diagnosing carotid stenosis by doppler sonography: state of the art J. Ultrasound Med. 24 1127–36 Grisan E, Pesce A, Giani A, Foracchia M and Ruggeri A 2004 A new tracking system for the robust extraction of retinal vessel structure Proc. 26th Annu. Int. Conf. IEEE-EMBS (San Francisco, CA, 1–4 September 2004) pp 1620–3 Hemachandra A H, Howards P P, Furth S L and Klebanoff M A 2007 Birth weight, postnatal growth and risk for high blood pressure at 7 years of age: results from the Collaborative Perinatal Project Pediatrics 119 e1264–70 Jarvisalo  M J, Jartti  L, Nanto-Salonen  K, Irjala  K, Ronnemaa  T, Hartiala  J J, Celermajer  D S and Raitakari O T 2001 Increased aortic intima-media thickness: a marker of preclinical atherosclerosis in high-risk children Circulation 104 2943–7 Jouret B, Dulac Y, Bassil R, Taktak A, Cristini C, Lounis N, Molinas C, Salles J P, Arnaud C, Acar P and Tauber M 2011 Endothelial function and mechanical arterial properties in children born small for gestational age: comparison with obese children Horm. Res. Paediatr. 76 240–7 Koklu E, Kurtoglu S, Akcakus M, Yikilmaz A, Coskun A and Gunes T 2007 Intima-media thickness of the abdominal aorta of neonates with different gestational ages J. Clin. Ultrasound 35 491–7 Langley-Evans S C 2009 Nutritional programming of disease: unravelling the mechanism J. Anatomy 215 36–51 Litwin M and Niemirska A 2009 Intima-media thickness measurements in children with cardiovascular risk factors Pediatr. Nephrol. 24 707–19 McGill H C Jr et al 2000 Effect of coronary heart disease risk factors on atherosclerosis of selected regions of the aorta and right coronary artery Ateriscler. Thromb. Vasc. Biol. 20 836–45 McMillen I C and Robinson J S 2005 Developmental origins of the metabolic syndrome: prediction, plasticity and programming Phys. Rev. 85 571–633 Molinari  F, Zeng  G and Suri  J S 2010 A state of the art review on intima-media thickness (IMT) measurement and wall segmentation techniques for carotid ultrasound Comput. Methods Prog. Biomed. 100 201–21 Simon A, Gariepy J, Chironi G, Megnien J L and Levenson J 2002 Intima–media thickness: a new tool for diagnosis and treatment of cardiovascular risk J. Hypertens. 20 159–69 Skilton M R, Evans N, Griffiths K A, Harmer J A and Celermajer D S 2005 Aortic wall thickness in newborns with intrauterine growth restriction Lancet 365 1484–6 Thompson  M, Dana  T, Bougatsos  C, Blazina  I and Norris  S L 2013 Screening for hypertension in children and adolescents to prevent cardiovascular disease Pediatrics 131 490–525 Touboul P J et al 2004 Mannheim intima–media thickness consensus Cerebrovasc. Dis. 18 346–9 van der Meer  I M, Bots  M L, Hofman  A, Del Sol  A I, van der Kuip  D A and Witteman  J C 2004 Predictive value of noninvasive measures of atherosclerosis for incident myocardial infarction: the rotterdam study Circulation 109 1089–94 Visentin S, Grisan E, Zanardo V, Trevisanuto D, Veronese E, Cavallin F and Cosmi E 2013 Developmental programming of cardiovascular risk in intrauterine growth restricted twin fetuses and aorta intima thickness J. Ultrasound Med. 32 279–84 Zanardo  V, Fanelli  T, Weiner  G, Fanos  V, Zaninotto  M, Visentin  S, Cavallin  F, Trevisanuto  D and Cosmi E 2011 Intrauterine growth restriction is associated with persistent aortic wall thickening and glomerular proteinuria during infancy Kidney Int. 80 119–23

6371