A reference free approach for the comparative evaluation of eight segmentation methods for the estimation of the left ventricular ejection fraction in cardiac MRI

45322 A reference free approach for the comparative evaluation of eight segmentation methods for the estimation of the left ventricular ejection fraction in cardiac MRI. Category: Scientific Session Communications Topic: Preclinical Studies and Basic Science / Processing and quantification: imaging Authors: A. Lalande, J. Lebenberg , I. Buvat, P. Clarysse, C. Casta, A. Cochet, C. Constantinidès , J. Cousty, A. de Cesare, S. Jehan-Besson, M. Lefort, L. Najman, E. Roullot, L. Sarry, C. Tilmant, M. Garreau, F. Frouin; Dijon,FR,University of Burgundy,LE2I (UMR CNRS 6306), Paris,FR,Inserm/UPMC,Laboratoire d'Imagerie Fonctionnelle, Ivry Sur Seine,FR,ESME-Sudria,PRIAM, Orsay,FR,Université de Paris Sud,IMNC (CNRS UMR 8165), Lyon,FR,Université de Lyon,CREATIS (CNRS UMR 5220), Marne La Vallée,FR,Université de Paris Est,LIGM-A3SI (UMR 8049), Aubière,FR,LIMOS,LIMOS (CNRS UMR 6158), Clermont-Ferrand,FR,Université d'Auvergne,ISIT (CNRS UMR 6284), Clermont-Ferrand,FR,Université Blaise Pascal,Institut Pasteur (CNRS UMR 6602), Rennes,FR,Université Rennes 1,LTSI (INSERM UMR 1099) Purpose / Introduction Objective evaluation and comparison of segmentation algorithms for medical imaging is still a challenging issue. The most frequently used evaluation method consists in comparing the segmentation with a manual delineation. Since obtaining such manual segmentation can be tedious, we proposed a method based on the ”extended Regression Without Truth” approach (eRWT)(1). This approach is applied to the comparative evaluation of 8 segmentation algorithms with different degrees of automation from the estimated left ventricular ejection fraction (LVEF). Subjects and Methods Series of short-axis cine MR images covering the left ventricle were obtained in 36 patients and 9 healthy volunteers (20 frames per cycle). For each series, five automated segmentation methods(2-6) with different degrees of automation were used, and manual contouring was also performed by three operators, providing 8 independent estimates of the LVEF (eLVEF). The principle of our non-supervised method consists in estimating the parameters of a linear model linking the measurements (eLVEF) and the true value (tLVEF). The distribution of the tLVEF is assumed to follow a Beta distribution (mu=4 and nu=5) : The relationship between the tLVEF and the eLVEF is given by : The determination of a, b and the standard deviation of the error makes it possible to compare the accuracy of the methods. Methods are ranked as a function of their accuracy, based on a figure of merit (Fm) : Results The obtained ranking suggests that the LVEF estimations provided by manual methods were the most accurate and the least variable. Methods Fm Manual 0.002 Manual 0.004 Manual 0.004 Semi-automatic(2) 0.006 Semi-automatic(3) 0.009 Automatic(4) 0.010 Semi-automatic(5) 0.011 Semi-automatic(6) 0.012 Discussion/Conclusion http://abstracts.webges.com/submission/preview/print_publication.php... 1 sur 2 12/02/2013 11:50 All automated methods considered in our study were less accurate than manual delineations. The main limitations aredue to segmentation failures in some particular slices. Their combination to provide one optimized mutual contour iscurrently under study.The eRWT approach proved to be relevant to compare different segmentation methods without the use of goldstandard and without any prior concerning the automation degree of the method. Moreover, it can be used to evaluatethe improvement of a method in progress. References(1) Lebenberg J. et al, 2012, IEEE Trans Med Imaging, in press (2) Constantinides C. et al, 2009, The Midas Journal(3) Lalande A. et al, 2004, JCMR, 817-827(4) Constantinides C. et al, 2012, IEEE EMBC, in press(5) Schaerer J. et al, 2010, Med Image Anal, 738-749(6) Cousty J. et al, 2010, Image Vision Comput, 1229-1243http://abstracts.webges.com/submission/preview/print_publication.php... 2 sur 212/02/2013 11:50