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

Area of Science:

• Medical imaging analysis
• Deep learning for medical image segmentation
• Cardiovascular disease diagnostics

Background:

• Accurate carotid intima-media thickness (CIMT) measurement is crucial for cardiovascular risk assessment.
• Automated CIMT analysis using deep learning can improve efficiency and consistency.
• U-Net and its variants are promising architectures for medical image segmentation tasks.

Purpose of the Study:

• To systematically compare the performance of U-Net and its variants for automated CIMT segmentation in ultrasound images.
• To evaluate both segmentation accuracy and real-time efficiency of different U-Net models.
• To provide guidance for selecting appropriate models based on clinical needs.

Main Methods:

• Ten U-Net-based models were trained and evaluated on the publicly available Carotid Ultrasound Boundary Study (CUBS) dataset.
• Standard preprocessing techniques and identical hyperparameters were used for all models.
• Performance was assessed using segmentation metrics (Dice, IoU, Precision, Recall, Accuracy) and efficiency metrics (training/inference time, parameter count).

Main Results:

• All models demonstrated high segmentation accuracy (Dice/IoU > 0.80/0.67).
• Attention U-Net achieved the highest segmentation accuracy.
• UNeXt exhibited the fastest training and inference speeds with a lower parameter count.
• UNet++ showed superior performance in generating smooth and natural boundaries.

Conclusions:

• U-Net-based models offer a viable solution for automated CIMT segmentation with high accuracy and efficiency.
• The choice of model involves a trade-off between segmentation accuracy and real-time performance.
• This study provides valuable insights for selecting U-Net variants tailored to specific clinical requirements in CIMT analysis.