Umeå universitets logga

Objective: Detection of atherosclerotic plaque in the carotid arteries is essential for early cardiovascular risk assessment. While B-mode ultrasound is the standard imaging modality, most deep learning studies focus on longitudinal views of the common carotid artery, with limited investigation of cross-sectional imaging and other clinically relevant regions such as the carotid bulb and split. To the best of our knowledge, this study presents the first region-specific evaluation of plaque segmentation using deep learning models in cross-sectional carotid ultrasound images.

Methods: Cross-sectional carotid ultrasound images from the 6 y follow-up of the population-based VIPVIZA cohort were analyzed. Two segmentation configurations were investigated: Case 1, where plaque segmentation was performed directly on the full field of view, and Case 2, where segmentation was guided by artery-centered regions of interest (ROIs) derived from manual ground truth annotations. Six deep learning architectures were evaluated: YOLOv8, U-Net, Attention U-Net, U-Net++, Swin U-Net and DeepLabV3. All images were pre-processed using intensity normalization and contrast-limited adaptive histogram equalization to enhance boundary visibility.

Results: In the full field of view configuration (Case 1), DeepLabV3 achieved the highest overall Dice score (0.65), followed by Swin U-Net (Dice = 0.64) and U-Net++ (Dice = 0.62), while YOLOv8 and Attention U-Net showed comparable performance (Dice = 0.61). In the ROI-based configuration (Case 2), U-Net achieved the highest segmentation performance in the common carotid artery (Dice = 0.73), while YOLOv8 achieved the best performance in the carotid bulb (Dice = 0.66). In the carotid split region, U-Net++ achieved the highest Dice score (0.87), followed closely by DeepLabV3 (Dice = 0.86) and Attention U-Net (Dice = 0.84).

Conclusion: Region-specific results from this study support the development of automated plaque segmentation in cross-sectional carotid ultrasound imaging. The findings highlight the importance of anatomical guidance and demonstrate that ROI-based approaches substantially improve segmentation performance. DeepLabV3 and transformer-based models show strong potential in full-image settings, while U-Net-based architectures remain highly effective at ROI-based segmentation.