Learning Wall Segmentation in 3D Vessel Trees using Sparse Annotations

TL;DR

This work tackles the challenge of training 3D carotid artery wall segmentation with sparse clinical annotations by combining centerline-based cross-section sampling, an adversarially trained 2D segmentation network, and Poisson surface reconstruction to generate 3D pseudo-labels for training a 3D nnU-Net. It introduces a bifurcation-aware sampling strategy that probes cross-sections perpendicular to the bifurcation axis, improving segmentation in the complex bifurcation region. The approach achieves robust 3D segmentation across healthy and stenotic cases, with limitations in distal ECA and bifurcation regions and dependency on 2D segmentation quality. This methodology enables efficient 3D wall assessment and downstream biomarker extraction, facilitating improved evaluation of carotid artery stenosis in clinical practice.

Abstract

We propose a novel approach that uses sparse annotations from clinical studies to train a 3D segmentation of the carotid artery wall. We use a centerline annotation to sample perpendicular cross-sections of the carotid artery and use an adversarial 2D network to segment them. These annotations are then transformed into 3D pseudo-labels for training of a 3D convolutional neural network, circumventing the creation of manual 3D masks. For pseudo-label creation in the bifurcation area we propose the use of cross-sections perpendicular to the bifurcation axis and show that this enhances segmentation performance. Different sampling distances had a lesser impact. The proposed method allows for efficient training of 3D segmentation, offering potential improvements in the assessment of carotid artery stenosis and allowing the extraction of 3D biomarkers such as plaque volume.

Figures (7)

• Figure 1: (left) Centerlines were extracted and the vessel wall was annotated in eight 2D cross-sections per carotid artery in 3D BB-MRI images. (adapted from blausen2014medical) (middle) The centerline and an adversarial 2D network are used to create densely distributed contours of the carotid artery wall. (right) These contours are used to train a 3D nnU-Net, enabling automatic 3D segmentation of the carotid artery. This allows the extraction of the vessel wall thickness and the visualization of the carotid geometry.
• Figure 2: (a) Placement of sparse annotations as proposed by Strecker et al. strecker2020carotidstrecker2021carotid (b) Annotation of the inner (cyan) and outer (orange) ICA wall. The ECA is visible in the cross-section but not annotated.
• Figure 3: (a) Cross-section sampling perpendicular to the centerline. In the bifurcation area, this can, e.g., lead to a large cut through the ECA and CCA when sampling the ICA centerline. (b) The bifurcation axis is defined as the line between the centerline bifurcation and the mass center of the ECA point and the ICA point. (c) Cross-section sampling perpendicular to the bifurcation axis in the bifurcation area and perpendicular to the centerline in the ICA, CCA, and ECA. (d) Bifurcation cross-section with overlay of automatic ECA and ICA contours (e) contours after combining ICA and ECA segmentation.
• Figure 4: (a) Cross-section of a T1-weighted MRI with ICA and ECA visible (b) Cross-section of the automatic 3D segmentation with ECA and ICA segmentation (c) Automatic ICA segmentation and expert ICA contour annotation.
• Figure 5: Incorrect segmentations in the ECA (Plane 8). On the left, the ECA segmentation does not continue up to the cross-section position. On the right, only parts of the ECA are segmented.