Automatic delineation of myocardial contours in late-enhancement long-axis cardiac MR images

Introduction

Viability assessment is essential for therapy planning following a myocardial infarction. In particular, the proportion of viable myocardium is a major factor in determining whether a patient may benefit from revascularization [1]. In addition to estimating the left ventricular myocardial thickness and thickening with functional imaging, it is possible to visualize normal and non-viable areas with high spatial resolution, using late-enhancement cardiac MR imaging (LECMR). To locate and quantify non-viable tissue, it is first necessary to delineate the endo- and epicardial contours on every available view of the LECMR acquisition. In particular long-axis (LA) views are useful because they provide a visualization of the apical area, which is often not well visible on or not covered well by short-axis slices. While manual delineation is a tedious and time-consuming task, its automation is challenging and, to our knowledge, not yet addressed by any publication or commercial product. This is mainly due to the non-homogeneous intensity of the myocardium resulting from contrast agent accumulation in infarcted areas.

Purpose

We propose a novel method to delineate the endo- and epicardial contours in late-enhancement long-axis cardiac MR images with a minimal user-input in order to provide an accurate quantitative viability assessment.

Methods

Before running the automatic segmentation process, the user selects one enhancement type from four pre-defined ones (no enhancement, diffuse/small enhancement, large/transmural scar, sub-endocardial enhancement), depending on his own observation of the LA view.

Then the delineation of the myocardial contours is performed by alternating automatic deformation of a geometrical template and computation of a binary map of the enhanced areas. The template is ribbon-shaped with a variable width, its position is updated depending on image gray values. The map is a 2D binary image showing enhanced areas, it is updated by thresholding the image gray values in a region of interest centered on the endocardium, in order to include sub-endocardial scars. The segmentation is performed as follows: (1) automatic delineation of myocardial contours on short-axis LECMR slices [2], (2) initialization of the geometrical template position and binary map based on step (1) results, (3) iterative loop between geometrical template deformation and update of enhanced areas map: each new position of the template leads to a new map computation, which is then used to deform the template again.

Results

The method was tested on 20 LA LECMR images acquired in a multi-center study between 2004–2007 (Philips Intera scanner 1.5 T, M FFE sequence, TE = 1.7 ms, TR = 4.5 ms, flip angle = 15°). All images are 256 × 256, with pixel size around 1.5 mm. Three experienced users manually delineated the myocardial contours to quantitatively assess the precision of the proposed method and evaluate inter-observer variability. The average error between the manual and automatic contours was 2.4 +/- 1.0 mm for the endocardium and 2.2 +/- 1.1 mm for the epicardium (see Table 1 for detailed results). The inter-observer variability was computed as an average distance from each manual contour to the other ones and was equal to 1.7 +/- 0.7 mm for the endocardium and 1.5 +/- 0.9 mm for the epicardium. As shown in Fig. 1, the visual quality is good, the contours successfully surround both normal and abnormal parts of the myocardium, which allows a reliable assessment of the percentage of non-viable tissue. Moreover, the accuracy (~1.5 pixel) is in the same range as inter-observer variability (>1 pixel). Remark: as accurate delineation of the valve plane is not required for viability assessment as long as myocardial contours are correct, it was not addressed in this study.

Table 1 Mean positioning error with respect to 3 manual contours and inter-observer variability. D_Refi is the mean distance to reference contour number i

Figure 1

Final myocardium contours on 4 different patients.

Conclusion

We presented a robust and efficient method for the automatic delineation of the myocardial contours in long-axis LECMR images.