We introduce a two-step segmentation process. First, segments the carotid artery to narrow the search space and focus on the region of interest around the artery. Then, it segments the calcified plaque within that targeted region. This approach achieves an intersection over union (IoU) of 0.9412 for the 2D model and 0.8095 for the 3D model, outperforming the baseline methods that directly segment plaques.
The dataset is from the Penn Stroke Registry, consisting of head and neck Computed Tomography Angiography (CTA) images, along with manually annotated segmentation files of calcified plaques, provided by medical professionals. There are three independent CTA neck imaging data sets: ESUS(N=70), CEA(N=27), and CAS(N=17). All cases had calcified carotid plaques manually segmented by two independent annotators and proofed by an expert radiologist.
First, using TotalSegmentator, we get the output of the common carotid artery left and right, vertebrae C3 and C5. Then we analyzed carotid artery intensity values and calculated a threshold based on the mean intensity. Then we set threshold ranges of -50 and +200. The images were then cropped around the carotid artery, guided by vertebrae segmentation from TotalSegmentator. Finally, we applied a region-growing technique to produce the carotid artery mask.
You can find the code here - Carotid Artery Segmentation
We segment plaque using 2D and 3D UNet. We conducted two different experiments.
-
ESUS dataset: For the 2D UNet, we achieved an IoU of 0.9412, while the 3D UNet resulted in an IoU of 0.8095. Both models segmented the calcified plaque effectively, but the 2D UNet demonstrated better performance.
-
ESUS + CEA + CAS: The 2D UNet achieved an IoU of 0.3920, whereas the 3D UNet yielded a significantly higher IoU of 0.6840. This suggests that the 3D UNet performs better across diverse datasets, indicating better generalization capability.
You can find the code here - 2D UNet, 3D UNet
2D UNet:
3D UNet:
This project is licensed under the MIT License - see the LICENSE file for details.
For any questions or comments, feel free to reach out to:
- Jenna Kim at [email protected]