Coronary artery segmentation in non-contrast calcium scoring CT images using deep learning

TL;DR

A deep learning algorithm for segmenting coronary arteries in multi-vendor ECG-gated non-contrast cardiac CT images which benefits from a novel framework for semi-automatic generation of Ground Truth (GT) via image registration is introduced.

Abstract

Precise localization of coronary arteries in Computed Tomography (CT) scans is critical from the perspective of medical assessment of coronary artery disease. Although various methods exist that offer high-quality segmentation of coronary arteries in cardiac contrast-enhanced CT scans, the potential of less invasive, non-contrast CT in this area is still not fully exploited. Since such fine anatomical structures are hardly visible in this type of medical images, the existing methods are characterized by high recall and low precision, and are used mainly for filtering of atherosclerotic plaques in the context of calcium scoring. In this paper, we address this research gap and introduce a deep learning algorithm for segmenting coronary arteries in multi-vendor ECG-gated non-contrast cardiac CT images which benefits from a novel framework for semi-automatic generation of Ground Truth (GT) via image registration. We hypothesize that the proposed GT generation process is much more efficient in this case than manual segmentation, since it allows for a fast generation of large volumes of diverse data, which leads to well-generalizing models. To investigate and thoroughly evaluate the segmentation quality based on such an approach, we propose a novel method for manual mesh-to-image registration, which is used to create our test-GT. The experimental study shows that the trained model has significantly higher accuracy than the GT used for training, and leads to the Dice and clDice metrics close to the interrater variability.

Figures (8)

• Figure 1: The challenge of segmenting the coronary vessel from non-contrast CT is demonstrated by an example of the Left Anterior Descending (LAD) artery visualized in Stretched Curvilinear Reformulation (CPR)---the same vessel is shown in non-contrast and contrast CT. In each case, we show the CT image alone (left) and an overlaid ML segmentation mask (right). The red arrows point to the areas where Hounsfield values of the vessel and surrounding tissue are practically the same, which makes applying classic segmentation techniques, like region growth methods, impossible. Note that the Hounsfield range for non-contrast and contrast CT is different, and is set to $(-120,200)$ and $(-120,800)$, respectively.
• Figure 2: Framework for semi-automatic GT generation and training of non-contrast coronary artery segmentation models.
• Figure 3: High-resolution LCA tree model with its skeletal representation in Blender. The tips and roots of individual Bones coincide with the vertices of the centerlines.
• Figure 4: The alignment of a 3D coronary vessel tree model (panels (a), (c), and (e)) with the non-contrast CT scan (panels (b), (d), and (f)) based on rotations of subsequent Bones. The black solid arrows and circles in the 3D views (a), (c), and (e) indicate the center of rotation at a given alignment step. The rotation directions are shown using dashed curves. The corresponding contours of the LCA tree at a fixed slice of the CT scan are depicted in the axial views (b), (d), and (f).
• Figure 5: Example of a mismatch between the GT created based on registration ($\mdblksquare$), the inference of the neural network ($\mdblksquare$), and the manually aligned contrast GT ($\mdblksquare$) in the distal part of RCA (a), juxtaposed with the corresponding non-contrast CT scan. Panel (b) is a CPR projection along the vessel centerline indicated in (a) by a solid black rectangle. Panels (c) and (d) show cuts in the vessel's short axis at two distinct points. Based on the visual comparison, we can note that the inference of the deep model and manually aligned vessel are consistent with the image, unlike the GT used in the training process.