Development and validation of intracranial thrombus segmentation on CT angiography in patients with acute ischemic stroke

Abstract

Background and purpose: Thrombus characterization is increasingly considered important in predicting treatment success for patients with acute ischemic stroke. The lack of intensity contrast between thrombus and surrounding tissue in CT images makes manual delineation a difficult and time consuming task. Our aim was to develop an automated method for thrombus measurement on CT angiography and validate it against manual delineation.

Materials and methods: Automated thrombus segmentation was achieved using image intensity and a vascular shape prior derived from the segmentation of the contralateral artery. In 53 patients with acute ischemic stroke due to proximal intracranial arterial occlusion, automated length and volume measurements were performed. Accuracy was assessed by comparison with inter-observer variation of manual delineations using intraclass correlation coefficients and Bland-Altman analyses.

Results: The automated method successfully segmented the thrombus for all 53 patients. The intraclass correlation of automated and manual length and volume measurements were 0.89 and 0.84. Bland-Altman analyses yielded a bias (limits of agreement) of -0.4 (-8.8, 7.7) mm and 8 (-126, 141) mm3 for length and volume, respectively. This was comparable to the best interobserver agreement, with an intraclass correlation coefficients of 0.90 and 0.85 and a bias (limits of agreement) of -0.1 (-11.2, 10.9) mm and -17 (-216, 185) mm3.

Conclusions: The method facilitates automated thrombus segmentation for accurate length and volume measurements, is relatively fast and requires minimal user input, while being insensitive to high hematocrit levels and vascular calcifications. Furthermore, it has the potential to assess thrombus characteristics of low-density thrombi.

Figure 1. Illustration of the pipeline of the automated method.

(A) The segmentation of the contralateral vasculature is started after manual seeds point placement. The lumen is segmented using a graph cut segmentation.(B) The thrombus is detected by mirroring the centerline and a Bspline registration followed by the segmentation of the occluded segment. Distal and proximal thrombus position is detected using the lumen radii.(C) Thrombus is segmented using tubular masking and region growing.

Figure 2. Illustration of the manual segmentation process.

A centerline is drawn in orthogonal views and subsequently corrected using MPR viewing. The contours of the thrombus are drawn first in a longitudinal view and subsequently corrected using on a transversal view.

Figure 3. Scatter plot of thrombus length and volume of observer 1 and automated method (A, B).

Scatter plot of thrombus length and volume of observer1 and observer 2 and 3 (C, D). The solid line represents the identity line.

Figure 4. Bland-Altman plot comparing thrombus volume and length as measured by observer 1 and the automated method (A, B), and for all observers (D, C).