High-accuracy Myocardial Detection by Combining Level Set Method and 3D NURBS Approximation

Abstract

Accurate detection of myocardium is much important in diagnosis of heart diseases. In this study, we proposed the myocardial detection method by combining level set method in 2D image and 3D NURBS approximation. We first extracted epi- and endocardial walls by level set method in 2D image. Calculation cost was reduced by the processing on the slice. In this extraction, we used a near-circular shape of left ventricle and set the initial circle in the myocardial region surrounding the endocardium. We then approximated these extracted walls with 3D NURBS (Non-uniformed rational Bspline) model. Here, we used third-order B-spline basis function. In our study, we applied the method to MRI T1-weighted heart images of 10 subjects (5 normal subjects and 5 patients with apical hypertrophic cardiomyopathy). The pixel size was 1.62 x 1.62 mm, and the slice interval and the number of slices were 6.62 mm and 18, respectively. Evaluation of our method was done by comparing with manual detection by two cardiologists. As a result, the accuracy of endocardial detection was about the same as or less than the difference between cardiologists. While on the other hand, that of epicardial detection was larger than the difference between cardiologists. We inferred that epicaridial contour is clearer than endocardial one. Average of detection error by the method combining level set method and NURBS approximation (endocardium: 2.58 mm / epicardium: 2.71 mm) was smaller or almost same as only level set method (2.77 mm / 2.51 mm). However, as for variance of the error, combining method (0.58 mm / 0.59 mm) was smaller than only level set method (1.07 mm / 1.06 mm). These results show that NURBS approximation suppressed the variation of the detection accuracy.