VIEWS: 0 PAGES: 4 POSTED ON: 5/28/2013 Public Domain
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. 250 All Points Filtered Points 200 unity Best Fit for Filtered Points Automatic Volumes 150 100 50 y = 0.9009x + 19.307 2 R = 0.7989 0 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 . 180 160 160 140 140 120 120 100 100 Volume Volume 80 80 60 60 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. References 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.