Improving the reproducibility of MR-derived left ventricular volume and function measurements with a semi-automatic threshold-based segmentation algorithm

Abstract

To validate a novel semi-automatic segmentation algorithm for MR-derived volume and function measurements by comparing it with the standard method of manual contour tracing. The new algorithms excludes papillary muscles and trabeculae from the blood pool, while the manual approach includes these objects in the blood pool. An epicardial contour served as input for both methods. Multiphase 2D steady-state free precession short axis images were acquired in 12 subjects with normal heart function and in a dynamic anthropomorphic heart phantom on a 1.5T MR system. In the heart phantom, manually and semi-automatically measured cardiac parameters were compared to the true end-diastolic volume (EDV), end-systolic volume (ESV) and ejection fraction (EF). In the subjects, the semi-automatic method was compared to manual contouring in terms of difference in measured EDV, ESV, EF and myocardial volume (MV). For all measures, intra- and inter-observer agreement was determined. In the heart phantom, EDV and ESV were underestimated for both the semi-automatic. As the papillary muscles were excluded from the blood pool with the semi-automatic method, EDV and ESV were approximately 20 ml lower in the patients, whereas EF was approximately 16 % higher. Intra- and inter-observer agreement was overall improved with the semi-automatic method compared to the manual method. Correlation between manual and semi-automatic measurements was high (EDV: R = 0.99, ESV: R = 0.96; EF: R = 0.80, MV: R = 0.99). The semi-automatic method could exclude endoluminal muscular structures from the blood volume with significantly improved intra- and inter-observer variabilities in cardiac function measurements compared to the conventional, manual method, which includes endoluminal structures in the blood volume.

Appendix

The semi-automatic segmentation method (QMass MR research edition, Medis, Leiden, The Netherlands) was based on the signal intensity distribution of MR images. The technique of normalized convolution [18] is used to estimate the spatially varying blood and muscle intensities within a user-provided epicardial contour. The voxel intensity is defined by a first order model with six variables:

$$ I(x,y) = (1 - w(x,y)) \cdot I_{m} (x,y) + w(x,y) \cdot I_{b} (x,y) $$

(3)

with:

$$ \begin{gathered} Muscle{:}\,I_{m} (x,y) = a_{0} + a_{1} x + a_{2} y \hfill \\ Blood{:}\,I_{b} (x,y) = b_{0} + b_{1} x + b_{2} y \hfill \\ \end{gathered} $$

(4)

where a₀, a₁, a₂, b₀, b₁ and b₂ are constants that vary among scans according to the grey value distribution of the image within the epicardial contour. I _m(x, y) and I _b(x, y) represent the approximation of the intensity of muscle and blood, respectively, at the position (x, y) (see Fig. 4). The constants are obtained with an iterative optimization procedure, during which the weight w(x, y) is initialized to either 1 or 0 using the Otsu threshold method [19].The procedure is stopped when the classification w > 0.5 is unaltered between iterations or when the number of iterations exceeds ten.

Assuming a linear relationship between the fraction of blood and the intensity of the voxel (I(x, y)), the weight w(x, y) represents the fraction of blood in the voxel.

$$ w(x,y) = \frac{{I(x,y) - I_{m} (x,y)}}{{I_{b} (x - y) - I_{m} (x,y)}} $$

(5)

Fig. 4

Segmentation procedure. a Shows a short axis view of the heart. The signal intensities along the red line in a are represented in b. The algorithm fits two planes (represented by the solid black lines) through the highest (blood) and lowest (muscle) signal intensities within the epicardium. A threshold plane was defined at 70 % between the two intensity planes. Voxels with signal intensities above this threshold are considered pure blood, and voxels with signal intensities below this threshold are considered to be pure muscle

A binary classification is obtained by thresholding w(x, y). If w(x, y) is higher than the threshold value, the voxel is considered pure blood, otherwise it is defined as pure muscle. In this experiment the threshold value was set to 70 %. Blood volume measures are obtained by multiplying the number of voxels classified as blood with the voxel volume.

Note that this method does not distinct trabeculations from papillary muscle and as such stands out from methods that do specifically target the papillary muscles (See e.g., [20]).