Accurate measurement of the carotid artery vessel wall thickness is paramount for monitoring patients with atherosclerosis. In 3D MRI, achieving precise vessel wall segmentation is essential, but segmenting the entire image is prone to error and time-consuming. This paper presents a fully automated deep learning-based segmentation framework that offers a practical and accurate solution for measuring carotid artery vessel wall thickness in MRI, potentially improving the monitoring and treatment of patients with atherosclerosis. We propose a multilevel EfficientNet-UNet++ architecture operating on a slice-by-slice basis and including contextual information. The approach entails four fundamental steps at different levels: (1) extracting the non-black region from the MRI, (2) identifying the region of interest (ROI) using the local 3D context, (3) upscaling the area around the segmented ROI, focusing on the vessel wall and surrounding tissue, and (4) re-segmenting the upscaled ROI to extract finer details of the vessel. Our multilevel segmentation approach outperforms state-of-the-art results in a public dataset comprising 2718 annotated slices from 75 patients, achieving a DICE similarity coefficient of 0.8704 on the COSMOS 2022 test set, surpassing nnU-Net (0.8527), the challenge-winning label-propagation method (0.8563), and DBF-UNet (0.8608). Experimentation and analysis demonstrate its effectiveness and potential value in clinical applications, with precise results near arterial bifurcation and in the presence of atherosclerotic plaque. The proposed method provides a reliable and accurate solution by focusing on the region of interest, providing local context, and increasing the input resolution, achieving competitive results with significant potential in clinical applications. • Fully automated deep learning approach for accurate vessel wall segmentation in 3D carotid artery MRI. • Multilevel strategy for region-of-interest identification and upscaled-resolution segmentation to capture fine details. • Improved segmentation accuracy near arterial bifurcations using local contextual information. • Improvement of the state-of-the-art performance and clinical potential demonstrated on a public dataset.
Gago et al. (Wed,) studied this question.