Perfusion weighted color maps for enhanced visualization of myocardial infarction by MSCT: preliminary experience

Abstract

Purpose The reliable detection of myocardial perfusion defects and myocardial infarction (MI) is of great interest in the comprehensive workup of coronary artery disease. The aim of this study was to optimize the ability of contrast-enhanced cardiac multislice spiral computed tomography (MSCT) for detecting hypoperfused myocardium as surrogate marker of MI using a newly developed post-processing technique. Methods First a model-based software tool for semi-automated detection of the long axis of the left ventricle and assignment of left-ventricular segments was developed using a region growing algorithm and a point distribution model. To visualize changes of the myocardial contrast enhancement pattern color coding was performed after spreading of the attenuation values. 15 patients (12 men, mean age 57 ± 15 years) with a history of MI underwent cardiac MSCT (16 × 0.75 mm, 120 kV, 550 mA s_eff., 100 ml Iopromide) and contrast enhanced delayed enhanced magnetic resonance imaging (DE-MRI) after administration of 0.2 mmol Gd-DTPA/kg/bodyweight as reference standard. Presence of infarction was assessed from MSCT, post-processed MSCT images and DE-MRI using a 17-segment model of the left ventricle. Results On DE-MRI MI was present in 78/255 myocardial segments. From conventional MSCT images MI was detected in 63/255 segments (5 false positive; sensitivity 74.4%; specificity: 97.1%), while on post-processed images MI was assigned to 74/255 segments (6 false positive; sensitivity 87.2%; specificity: 96.6%). Agreement between DE-MRI and conventional MSCT images for detecting MI was κ = 0.756. Using post-processed images agreement improved to κ = 0.850. Conclusion MSCT detection of hypoperfused myocardium as surrogate for MI can be improved using dedicated post processing algorithms.

Appendix

In the described approach a statistical segmentation method is applied that keeps track of learned shapes of the epi- and endocardium surfaces to avoid anatomically incorrect segmentation results and which is based on a Point Distribution Model (PDM) [19, 20]. Herein, a training set of patient-individual surfaces is defined by expert hand which is the basis for our segmentation technique. More specifically, each trained surface is sampled by regularly distributed n surface points and then a combined mean model ${\bar {\mathbf x}},$ representing an average epi- and endocardial point cloud, is calculated (Fig. 1a). By the application of a Principle Component Analysis (PCA) [19] of the training surfaces, we are able to get the covariance matrix S and its Eigenvectors in order to mathematically describe the different left ventricular learned shapes. Using this, one is able to approximate each trained dataset x by means of the average left ventricular model ${\bar {\mathbf x}}$ by the following equation in a very intuitive mathematically description:

$$ {\mathbf{x}} = {{\bar {\mathbf x}}} + \Phi {\mathbf{b}} $$

(1)

The matrix Φ contains the Eigenvectors φ _i of the t-largest Eigenvalues λ_i of the covariance matrix S. The vector b is the model parameter vector that weights each mode such that

$$ {\mathbf{b}}\, = \Phi ^T \left( {{\mathbf{x}} - {{\bar {\mathbf x}}}} \right). $$

(2)

In order to enforce the limitation to plausible shapes, often a hard limit of $\left| {b_i } \right| \le 3\sqrt {\lambda _i }$ is applied.

The segmentation of the myocardium starts by a single user click in the ascending aorta. Then, a region-growing algorithm is started that identifies the blood-pool of the left ventricle, which is the basis for the PDM model described above. By the application of a PCA of the blood-pool voxels, the main axis of the left ventricle is computed. Moreover, by means of morphological operators [20], one can fill the papillary muscles within the segmentation result and therefore extract the endocardial surface by the use of a snake technique which leads to very robust results in terms of the endocardium [21]. Since the basis of this approach is a technique that is based on anatomical statistics, we can apply the extracted endocardium surface to bring the algorithm (PDM) close to its final epicardial border. In order to get the final and correct anatomical epicardium surface, one has to detect the border within the image by robust image processing algorithms. For this purpose a combination of both, the Mahalanobis Distance [20] and the Mean-Shift technique is applied [22]. Furthermore, we introduce an expected grey value at the sample position which has been extracted from the shape-learning procedure which leads to the following weighted expression:

$$ f(x) = w_{{\text{ma}}} \cdot {\text{Mahalanobis}}(x) + w_{{\text{me}}} \cdot {\text{Meanshift}}(x) + w_{{\text{ex}}} \cdot {\text{ExpectedGrey}}(x) $$

(3)

The equation above has to be maximized for each of the sampling rays along the surface normal. The parameters w _ma, w _me and w _ex are weights in order to be able to emphasize each of the functional elements of Eq. 3.

Finally, in order to initialize the PDM procedure, including Eq. 3, and to determine the sample points along the epicardium surface, one has to introduce a LV-specific sampling scheme. Herein, at mid-ventricular position, we use parallel sampling, orthogonal to the LV long-axis. At apical and basal position, a half-spherical sampling shape is applied (Fig. 1b). In consequence, sampling rays, which apply Eq. 3, are defined in circular order, rotating around the LV long-axis. The length of the long-axis is defined by the region-growing in combination with the PCA of the blood-pool voxels. The Eigenvectors can then be used for correct scaling of the LV to the mean shape.

After extraction of the myocardium from the MSCT data set a 17 segment model of the left ventricle is automatically assigned to the segmentation results [12]. Thereafter a dedicated visual improvement technique was introduced in order to enhance the visual perception of possibly infarcted areas. For this purpose the intensity interval of the myocardium mass voxels has to be spread. For the described post-processing technique an empiric range from 50 HU to 140 HU that was considered to be representative for diagnostic results was used [4, 23]. This interval is stretched up to 1000 units including a linear interpolation of the voxels. Intensities below or above are not taken into account.

Finally, the selected voxels were colored using a predefined lookup table ranging from blue via green to yellow and red and which only takes care of the relative grey values of the myocardial voxels. The lower the perfusion of the myocardial tissue, the lower is the intensity in the original image. Such tissue is colored in blue. In the opposite direction, the better the perfusion, the brighter are the voxel intensities. In this direction the scale ends with yellow.