Performance Analysis of Deep Learning Models for Segmentation of Carotid Artery Vessel Wall in 3D-MERGE Images

Carotid vessel wall segmentation and determination of the lumen area are crucial for the diagnosis of atherosclerosis. U-Net-based deep learning models have been investigated for carotid vessel wall segmentation in magnetic resonance imaging. However, the use of these deep learning models for 3D Motion-Sensitized Driven Equilibrium-prepared Rapid Gradient Echo (3D-MERGE) imaging is less explored. In addition, the effect of preprocessing techniques on the performance of deep learning models using 3D-MERGE images need to be investigated.

This paper explores deep learning-based image segmentation models for carotid artery vessel wall segmentation from 3D-MERGE images. A detailed comparative analysis of U-Net, Attention U-Net, and Residual U-Net models with different preprocessing techniques is performed on a public dataset. The efficiency of the models is analyzed using various evaluation metrics including Dice score, sensitivity, and specificity.

The U-Net model achieved a Dice score of 70.85%, while the Attention U-Net gave 67.04%, showing a significant improvement ( p <0.05). Cross-validation analysis showed improved performance of the Attention U-Net model. Preprocessing reduced the number of erroneous detections by approximately 52% in both U-Net and Attention U-Net models. Among the three models, the Residual U-Net model underperformed in the segmentation of carotid vessel walls.

Our findings show that the U-Net and Attention U-Net models have great potential for detecting carotid vessels in 3D-MERGE images. Image preprocessing has a notable impact on the training of U-Net-based models.