Automatic Coronary Wall and Atherosclerotic Plaque Segmentation from 3D Coronary CT Angiography

Coronary plaque burden measured by coronary computerized tomography angiography (CCTA), independent of stenosis, is a significant independent predictor of coronary heart disease (CHD) events and mortality. Hence, it is essential to develop comprehensive CCTA plaque quantification beyond existing subjective plaque volume or stenosis scoring methods. The purpose of this study is to develop a framework for automated 3D segmentation of CCTA vessel wall and quantification of atherosclerotic plaque, independent of the amount of stenosis, along with overcoming challenges caused by poor contrast, motion artifacts, severe stenosis, and degradation of image quality. Vesselness, region growing, and two sequential level sets are employed for segmenting the inner and outer wall to prevent artifact-defective segmentation. Lumen and vessel boundaries are joined to create the coronary wall. Curved multiplanar reformation is used to straighten the segmented lumen and wall using lumen centerline. In-vivo evaluation included CCTA stenotic and non-stenotic plaques from 41 asymptomatic subjects with 122 plaques of different characteristics against the individual and consensus of expert readers. Results demonstrate that the framework segmentation performed robustly by providing a reliable working platform for accelerated, objective, and reproducible atherosclerotic plaque characterization beyond subjective assessment of stenosis; can be potentially applicable for monitoring response to therapy.

The modules and block diagram of the proposed framework were previously demonstrated in Figure   1 and further explained in expanded forms in the appendix Figures A for the segmentation modules and visualization modules in Figure B below. The implementation details of every modules are further described in the following paragraphs.

Data preparation module:
In this preparation step, input image format was converted to VTK and ITK images formats from Digital Imaging and Communications in Medicine (DICOM) image format.
VTK and VMTK toolkits use and generate the same image format, vtk::vtkImageData, while ITK uses its own data format, itk::Image. Two additional image format converters, vtk::vtkImageData to itk::Image and vice versa, were used during the transformation of the images among the toolkits to assure compatibility. To reduce the execution time, only the region of interest (ROI) was converted and processed by the framework. The 3D ROIs were automatically generated around the seed points.
The radiologist was able to manually adapt the ROIs if needed. Figure A, generating the lumen initial contour was performed using vesselness, region growing, and image intersection algorithms. The VMTK was utilized for implementing the Frangi's vesselness while ITK performs the connected threshold region growing algorithm and image intersection as listed in Table A.

Lumen initial contour module: As shown in
Feature image module: The gradient image was calculated using gradient magnitude with smoothing algorithm then the sigmoid function was applied. The generated feature image was then used for both lumen and vessel segmentation. This module was implemented using only ITK as listed in Table A.
Level Set segmentation module: The Geodesic level set algorithm was used for segmenting both lumen and vessel. The module inputs were the initial contour and the feature image. The vessel initial contour was calculated using the segmented lumen as described in the next paragraph. In lumen segmentation, the level set evolution parameters were set to extend the lumen initial contour to the lumen boundary. While in vessel segmentation, the parameters were set to shrink the initial contour toward to the vessel boundary. The level set segmentation was performed using VMTK.
Vessel initial contour module: The vessel initial contour was calculated by dilating the segmented lumen up to 5.5 mm to insure including the vessel and the plaques within the initial contour. In the segmented lumen image, the pixels within the lumen have negative values while the surrounding pixels have positive values. As a result, we dilated the lumen by applying morphological erosion. To get more accurate initial contour, the surrounding fat regions, identified as pixels with negative HU in the original image, were excluded.

3D mesh generation module:
This module was used during the visualization part for generating 3D meshes from the segmented images. The Marching Cube algorithm was applied on the segmented images to extract a zero-level set that was used by triangulation algorithm to generate an initial 3D mesh. The final 3D mesh was acquired by utilizing Laplacian smoothing algorithm. The implementation of this module was performed using VTK as listed in Table A