Performance Comparison of U-Net and Its Variants for Carotid Intima-Media Segmentation in Ultrasound Images.

Background/Objectives: This study systematically compared the performance of U-Net and variants for automatic analysis of carotid intima-media thickness (CIMT) in ultrasound images, focusing on segmentation accuracy and real-time efficiency. Methods: Ten models were trained and evaluated using a publicly available Carotid Ultrasound Boundary Study (CUBS) dataset (2176 images from 1088 subjects). Images were preprocessed using histogram-based smoothing and resized to a resolution of 256 × 256 pixels. Model training was conducted using identical hyperparameters (50 epochs, batch size 8, Adam optimizer with a learning rate of 1 × 10-4, and binary cross-entropy loss). Segmentation accuracy was assessed using Dice, Intersection over Union (IoU), Precision, Recall, and Accuracy metrics, while real-time performance was evaluated based on training/inference times and the model parameter counts. Results: All models achieved high accuracy, with Dice/IoU scores above 0.80/0.67. Attention U-Net achieved the highest segmentation accuracy, while UNeXt demonstrated the fastest training/inference speeds (approximately 420,000 parameters). Qualitatively, UNet++ produced smooth and natural boundaries, highlighting its strength in boundary reconstruction. Additionally, the relationship between the model parameter count and Dice performance was visualized to illustrate the tradeoff between accuracy and efficiency. Conclusions: This study provides a quantitative/qualitative evaluation of the accuracy, efficiency, and boundary reconstruction characteristics of U-Net-based models for CIMT segmentation, offering guidance for model selection according to clinical requirements (accuracy vs. real-time performance).