wright by iasiatube.news


									         Automated 3D Echocardiography Analysis Compared with
                    Manual Delineations and MUGA
                                                  G.J.T. Wright, J.A. Noble ∗
                Department of Engineering Science, University of Oxford, Parks Road, Oxford OX1 3PJ, UK.

         Abstract. In this paper we present a method for improving the volume results of a semi-automated 3D boundary
         tracking protocol similar to that reported in [1] (the only user interaction involves placement of an ellipsoid to
         initialize surface fitting). Specifically the refined approach uses an approximate shape model to filter the set of
         candidate boundary points and to reject points that arise due to noise, or undesirable features. We show how
         this leads to an improvement on the 3D echo analysis protocol. In particular we show how ejection fractions
         (EFs) calculated using the new protocol are within 5% of MUGA EFs and that the volume-time curves are
         close to those from manually drawn contours. To our knowledge this shows, for the first time, that an automatic
         echocardiographic analysis method can give clinically acceptable EFs.

1     Introduction

The recent introduction of commercial 3D mechanically scanned probes has generated considerable interest in
automated 3D echo analysis for example at Yale [2, 3], INRIA [4], and our own laboratory [1, 5].

In this paper we present a method for improving the volume results of a semi-automated 3D boundary tracking
protocol similar to that reported in [1] (a manually placed ellipsoid rather than the automatic cluster is used as the
initialization for the fitter). Specifically the refined approach uses an approximate shape model to filter the set of
candidate boundary points and to reject points that arise due to noise, or undesirable features. We show how this
leads to an improvement on the 3D echo analysis protocol. In particular we show how ejection fractions (EFs)
calculated using the new protocol are within 5% of MUGA EFs and that the volume-time curves are close to those
from manually drawn contours. To our knowledge this shows for the first time that an automatic echocardiographic
analysis method can give clinically acceptable EFs.

Other methods for reducing the number of spurious points have been reported by Stetten et al. [6], and suggested
in Sanchez-Ortiz et al. [1]. The former uses a symmetry measure that locates segmented points on either side of
the ventricle on real-time 3D ultrasound data sets. This would not work on the data used in our work as many of
the slices do not contain ample points on both sides of the ventricle. Further, the normals that we have are not the
true (3D) normals of the heart surface, but are projected onto the imaging plane. The latter calculates a region of
interest in each 2D temporal sequence directly from the image but has not been tested on real data to date.

The rest of this paper is presented as follows: in section 2 we outline the protocol for this work. We explain the
method of filtering in section 3. In section 4 we present the experimental results on 6 sets of clinical data, and in
section 5 we present our conclusions.

2     Tracking Protocol

The 3D analysis protocol has 5 main steps: image acquisition, segmentation, filtering, fitting and quantification.
Steps 1, 2, 4 and 5 are discussed below, with the filtering method (step 3) being discussed in section 3.

Image Acquisition. 14 patients with good acoustic windows for standard 2D B-mode echocardiography were
invited to participate in a study of 3D analysis methods in conjunction with the John Radcliffe Hospital, Oxford.
Six of the best of these were used in this study. Data sets of 12 evenly spaced co-axial long axis planes were
acquired over one cardiac cycle using a HP Sonos 5500 transducer in fundamental mode (3.5MHz and frame rate
of 25Hz). Although patients were selected with good acoustic windows for 2D, the quality of the images from the
3D probe are of considerably lower quality, and would be considered moderate to poor by clinicians if they were
standard 2D B-mode images (i.e. the endocardial border delineation can be quite poor), as reported in [7].

Segmentation. A 2D version of the FAIR segmentation algorithm [8,9] with fixed parameters was used to provide
a set of candidate boundary points for every slice at each time frame. This algorithm also supplies the local
    ∗ E-mail:   {gabriel,noble}@robots.ox.ac.uk
boundary normal associated with the candidate boundary points. Under ideal circumstances these point into the
LV cavity for the endocardial boundary.

Fitting. As in [1, 5], the ICP based fitting algorithm was used to fit a spline surface from one frame to the next.
In [1] the complete set of candidate boundary points produced by the FAIR algorithm was used and the fitter relied
upon to reject erroneous points. In this paper we augment this step as described in section 3.

Quantification The volumes of the fitted surfaces were calculated geometrically using standard mesh analysis.

3   Filtering Methodology

We assume that the left ventricle of the heart forms a closed surface (i.e. ignore the effect of the mitral value), and
in ultrasound images the LV cavity appears mainly dark while the myocardium appears mostly light. A filter based
on both the position of a boundary point and the direction of its normal was defined as follows:

At each time frame an ellipsoid was fitted to the complete set of candidate boundary points. This was used to find
an approximation to the axis of the left ventricle from the eigenvector corresponding to the largest eigenvalue. The
ellipsoid centre was used to define an approximation to the centre of the LV. Consider a point, p, that is a member
of the set of candidate boundary points and consider two lines. The first line, r, is the projection in the direction of
the point’s normal, np , passing through the point described by equation 1. The second, s, is the major axis of the
ventricle, defined by its centre, c, and the direction of the major axis, a, as in equation 2. Then,

                                                 r   =    p + κnp ,                                                (1)
                                                 s   =    c + λa,                                                  (2)

where κ and λ are scalars. Using a simple vector geometry argument the points of closest approach of the two
lines, r = R and s = S, are defined by,

                                                      (c − p) ∧ (a ∧ np )
                                        κR    = a·                   2    ,                                        (3)
                                                             a ∧ np
                                                         (c − p) ∧ (a ∧ np )
                                        λS    = np ·                   2       ,                                   (4)
                                                              a ∧ np

where ∧ denotes the cross product. The values of κR and λS can be used to filter out points based on the following
criteria. First we select points whose normals point into the ventricle. As a second constraint we take points which
result in S lying within the ventricle. Thirdly we filter points based on their distance from the axis by placing upper
and lower bounds on accepted values of κR . Mathematically, these three constraints can be written as:

                                              Dmin ≤ κR ≤         Dmax                                             (5)
                                                    |λS | ≤       Λ                                                (6)

where Λ is related to the length of the ventricle, while Dmin and Dmax are related to the width of the ventricle.
These values are estimated from the fitted ellipsoid for a given time frame and vary over the cardiac cycle.

4   Evaluation of Results

In this section the refined 3D analysis method is evaluated and compared with using unfiltered points. We consider
the volumes over the whole temporal sequence not just at end diastole and end systole. We also compute ejection
fractions and compare these to those from manual contours as well as MUGA.

Figure 1 shows a scatter plot of the volumes from the new filtering method and also unfiltered data against the
volumes of manually drawn contours.

It can be seen that overall the new method agrees better with the volumes of the manual contours than using
unfiltered points. Figure 2 shows sample volume time curves for two of the patients, showing curves from all three
methods. Using the manual volumes as the reference, it is clear that the new method has corrected errors that were
present with unfiltered points.

                                                                                                           All Points
                                                                                                           Filtered Points
                                                                                          200              unity
                                                                                                           Best Fit for Filtered Points

                                                                      Automatic Volumes



                                                                                                                y = 0.9009x + 19.307
                                                                                                                     R = 0.7989
                                                                                                0              20             40               60              80              100               120           140          160
                                                                                                                                                         Manual Volumes

Figure 1. Scatter graph of the calculated volumes against the volumes of the manually drawn contours for both
unfiltered and filtered data, volume is measured in cm3 .

                        180                                                                                                                                                   160





                                                                                                                                                                                                                                                         Manual Volume
                                                                                                         Manual Volume
                                                                                                                                                                               40                                                                        Unfiltered Volume
                         40                                                                              Unfiltered Volume
                                                                                                                                                                                                                                                         Filtered Volume
                                                                                                         Filtered Volume

                         20                                                                                                                                                    20

                          0                                                                                                                                                     0
                              1   2   3   4   5   6   7   8    9    10                     11       12    13        14   15        16     17        18                               1   2   3    4    5   6   7   8   9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
                                                              Time Frame                                                                                                                                                                 Time Frame

Figure 2. Examples of the volume-time curves produced by manual delineation and automatically using filtered
data and unfiltered data, volume is measured in cm3 . a) Patient 2, b) Patient 6.

4.1 Analysis 1. Table 1 shows the error in volumes calculated with respect to manually drawn contours for un-
filtered points and the new method over all the time frames of all the patients (n = 142). The mean, µ, standard
deviation, σ, and the RMS value of these errors are shown, all values are in cm3 .

                                         Data Type    µ        σ                                                                                                                                                   RMS
                                         Unfiltered 39.89 39.65                                                                                                                                                     56.15
                                           Filtered  9.36 12.20                                                                                                                                                    15.35
Table 1. Global errors in volumes over all patients and time frames.                                                                                                                                               This is based on six patient data sets leading
to a total of 142 points. All values are in cm3 .

It can be seen that the new filtering method results in a considerably lower mean error in volume, as well as a lower
standard deviation and RMS error value. However we can not draw absolute conclusions from this as manual
delineation are not necessarily a good reference, as is demonstrated in section 4.3.

4.2 Analysis 2. Table 2 compares the end diastolic volume, end systolic volume and the maximum absolute error
for the 3 methods. It can been seen that in each case the maximum absolute error between the new automatic
method volume and the manual derived volume are better than the equivalent measure for the method using the
unfiltered data. It can also be seen that in each case the volumes at end diastole and end systole are closer to the
manually derived volume with the new method than with unfiltered points.

4.3 Analysis 3. Table 3 shows the ejection fractions from each patient as calculated from MUGA, manual contours
and the two automatic methods. The ejection fractions from the new method agree better with the MUGA results
than either the manual ejection factions and the ejection fractions from unfiltered points. They are also within 5
percentage points of the MUGA results.

5   Conclusions and Future Work

We have shown how geometrically driven filtering of candidate boundary points improves the result of an auto-
mated fitting process, and have shown, using multiple data sets of various qualities, that the new method allows
for considerable improvement in the end result of the fitting process in terms of the derived volumes (and hence
the ejection fractions). Calculated ejection fractions for the new method are within 5% points of MUGA which is
clinically acceptable [10]. We did however take the best 6 of 14 data sets so this result will not be as good for lower
quality images. The current limitation of using this methodology in clinical practice is the transducer technology.
As this improves so we expect to see the modality produce good results for a larger range of patients.
                      End Diastolic Volume (cm3 )        End Systolic Volume (cm3 )      Max Error (cm3 )
Patient Frames Manual Unfiltered Filtered Manual Unfiltered Filtered Unfiltered Filtered
   1        31       133.76      209.70     140.31      77.35     140.97      102.20      97.90       35.91
   2        18       157.23       99.89     154.12     106.98      61.31      100.36      58.65       18.05
   3        25       113.54      203.98     146.94      76.89     150.80      98.04      106.25       41.18
   4        19        73.79      102.40      84.13      35.30      63.00      49.43       39.33       15.91
   5        20        89.35      124.46      92.11      48.23      75.29      49.59       41.66       13.83
   6        32       107.57      146.24     106.41      50.23      80.84      60.78       43.06        7.30
     Table 2. End diastolic volume end systolic volumes and the maximum absolute error for each patient.

                                         Ejection Fractions (%) (% error to MUGA)
                   Patient MUGA              Manual              Unfiltered        Filtered
                      1          27      42.17 (15.17) 32.78          (5.17)  27.16 (0.16)
                      2          38      31.96 (-6.04) 38.62          (0.62)  34.88 (-4.22)
                      3          35      32.28 (-2.72) 26.07 (-8.93) 33.28 (-1.72)
                      4          48      52.16 (4.16) 38.48 (-19.84) 50.73 (2.73)
                      5          54      46.02 (-7.98) 39.50 (-14.50) 51.15 (-3.85)
                      6          51      53.30 (2.30) 44.72 (-6.28) 52.80 (2.80)
Table 3. Ejection Fractions for each patient. Percentile errors to MUGA Ejection Fractions are shown in brackets

In our case absolute volumes could not be obtained from MUGA (as this would have required a calibration against
radioactive blood samples), and so we used the manual as a reference. In practice it represents a good estimation
to the true boundary. However the manual contour can be inaccurate as there are parts of the cardiac cycle and
regions of the heart where it is not possible to see the endocardial boundary completely. Future work will conduct
a similar study comparing the new 3D echo protocol with MRI.

Having shown that good volume matches can be produced by the 3D analysis protocol the next goal is to show
that good regional motion estimates can also be produced, although this would be considerably more reliable with
better quality images than currently provided by standard mechanically scanned 3D probes today.

Acknowledgements: The authors would like to thank Dr. Nigel Clark and Dr. Andrew Kellion of the John
Radcliffe Hospital, Oxford for supplying the data, and also clinical advice. GJTW gratefully acknowledges the
EPSRC for a postgraduate research studentship.

 1. G. I. Sanchez-Ortiz, J. Declerk, M. Mulet-Parada et al. “Automating 3D Echocardiographic Image Analysis.” In MICCIA,
    pp. 687–696. Pittsburgh, PA, USA, 2000.
 2. X. Papademetris. Estimation of 3D Left Ventricular Deformation from Medical Images Using Biomechanical Models.
    Ph.D. thesis, Yale University, May 2000.
 3. X. Papademetris, A. J. Sinusas, D. P. Dione et al. “Estimating 3D Strain from 4D Cine-MRI and Echocardiography:
    In-Vivo Validation.” In MICCAI, pp. 678–696. Pittsburgh, PA, USA, 2000.
 4. J. Montagnat & H. Delingette. “Space and Time Shape Constrained Defomable Surfaces for 4D Medical Image Segmen-
    tation.” In MICCIA, pp. 196–205. Pittsburgh, PA, USA, 2000.
 5. G. I. Sanchez-Ortiz, J. A. Noble, G. J. T. Wright et al. “Automating LV motion analysis from 3D echocardiology.” In
    Proceedings of Medical Image Understanding and Analysis, pp. 85–88. Oxford, UK., 1999.
 6. G. D. Stretten & S. M. Pizer. “Medial-Node Models to Identify and Measure Objects in Real-Time 3-D Echocardiography.”
    IEEE Transactions on Medical Imaging 18(10), pp. 1025–1034, October 1999.
 7. S. De Castro, L. Agati, D. Cartoni et al. “Harmonic imaging with levovist for transthoracic echocardiographic reconstruc-
    tion of left ventricle in patients with post-ischemic left ventricular function and suboptimal acoustic windows.” Journal of
    the American Society of Echocardiography 13(2), pp. 139–145, February 2000.
 8. M. Mulet-Parada & J. A. Noble. “2D+T Boundary Detection in Echocardiography.” In MICCAI, pp. 806–812. Cambridge,
    Mass. U.S.A., 1998.
 9. M. Mulet-Parada. Intensity independent feature extraction and tracking in echocardiographic sequences. Ph.D. thesis,
    Department of Engineering Science, Oxford University, 2000.
10. F. J. T. H. Wackers, H. J. Berger, D. E. Johnstone et al. “Multiple Gated Cardiac Blood Pool Imaging for Left Ventricular
    Ejection Fraction: Validation of the Technique and Assessment of Variability.” American Journal of Cardiology 43,
    pp. 1159–1166, June 1979.

To top