Improved Automated Quantification of Left Ventricular Size and Function from Cardiac Magnetic Resonance Images C Corsi 1,2 , F Veronesi 1 , C Lamberti 1 , V Mor-Avi 2 1 University of Bologna, Bologna, Italy 2 University of Chicago, Chicago, Illinois, USA Abstract Assessment of left ventricular (LV) size and function from cardiac magnetic resonance (CMR) images requires manual tracing of LV borders on multiple 2D slices, which is subjective, experience dependent, tedious and time-consuming. We tested a new method for automated dynamic segmentation of CMR images based on a modified region-based model, in which a level set function minimizes a functional containing information regarding the probability density distribution of the gray levels. Images (GE 1.5T FIESTA) obtained in 9 patients were analyzed to automatically detect LV endocardial boundaries and calculate LV volumes and ejection fraction (EF). These measurements were validated against manual tracing. The automated calculation of LV volumes and EF was completed in each patient in <3 min and resulted in high level of agreement with no significant bias and narrow limits of agreement with the reference technique. The proposed technique allows fast automated detection of endocardial boundaries as a basis for accurate quantification of LV size and function from CMR images. 1. Introduction It is widely agreed that comprehensive evaluation of cardiac function for the diagnosis and therapeutic follow up of myocardial pathologies requires a wide range of information. Thus, left ventricular (LV) volume over time curves provide clinically important information on LV dynamics, beyond the traditional ejection fraction (EF), which include direct insight into LV contraction and relaxation properties closely related to pathophysiology of various disease states. Cardiac Magnetic Resonance (CMR) provides noninvasive, high-resolution, radiation- free, dynamic imaging of the heart that allows accurate and reproducible evaluation of LV volumes throughout the cardiac cycle. Over the last decade, this methodology has become the standard reference technique for LV volume and EF measurements, against which other techniques are validated [1,2]. Although automated LV endocardial boundary detection is available in commercial software for analysis of CMR images, it is usually based on algorithm parameters that are sensitive to image quality and frequently depend on the specific imaging protocol [3]. Since optimization of these parameters for each individual pulse sequence is not possible, the computation of LV volumes and EF in clinical practice relies on frame-by-frame manual tracing of endocardial contours on multiple short-axis planes. This procedure is subjective, tedious, time-consuming and experience- dependent. Furthermore, its accuracy relies on geometrical models, such as disk-area summation, which may not always yield accurate results. Accordingly, our aim was to develop and test a technique for fast, automated, dynamic segmentation of CMR images, that would take into account image attributes specific to each pulse sequence. Our approach uses a region-based level set model described by Chan and Vese in [4]. This segmentation model is based on the minimization of an energy function containing information regarding the grey level values of the pixels into the image. The minimization of this energy function leads to the segmentation of the image in regions for which the difference in the grey level intensity average inside and outside is maximized. In our model we keep the region-based approach and embed in the segmentation model the a priori knowledge of statistical distribution of grey levels in CMR data: therefore the proposed method drives the curve evolution to achieve a maximum likelihood segmentation of the target with respect to the statistical distribution law of image pixels. We consider the noise in CMR images to have a Rician probability density function that approaches a Gaussian function when pixel intensity is higher than the noise level [5,6]. Our segmentation method was implemented in the 3D domain and requires a simple definition of a reference point of view within the data as initial condition for the dynamic detection of the LV endocardial boundaries throughout the cardiac cycle. ISSN 0276-6547 53 Computers in Cardiology 2006;33:53-56.