Automatic tuning of left ventricular segmentation of MR images using genetic algorithms

Cardiac magnetic resonance imaging provides an important tool for the quantitative analysis of left ventricular function. Automatic segmentation of left ventricular endocardium allows the evaluation of its functional performance. Since automatic segmentation algorithms are sensitive to the image characteristics, we designed a self-adaptive optimization system for an automated cardiac left ventricular contour detection algorithm. A Genetic Algorithm (GA) was used as a tuning method to optimize the settings of the automated segmentation of the MR Analytical Software System (MASS) package. The performance of the tuning method was evaluated on 20 clinically obtained short-axis examinations (10 using a TrueFisp and 10 using a Gradient echo pulse sequence) comparing manually and automatically detected contours. The degree of similarity, defined as the percentage of points that are similar between two contours, was used as a quantitative measurement of the performance of the segmentation algorithm. After optimization, the average degree of similarity between automatically detected and manually drawn endocardial contours increased from 58% to 70%. Compared with the inter-observer agreement of 73%, we conclude that GA-based optimization is an effective and efficient method to increase the reliability of our automated contour detection.