Joint 2D-3D cross-pseudo supervision for carotid vessel wall segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Research

Background:

• Accurate carotid vessel wall segmentation is vital for diagnosing atherosclerosis.
• Existing supervised learning methods struggle with limited or discontinuous 3D data.
• There is a need for robust segmentation techniques that overcome data limitations.

Purpose of the Study:

• To develop a novel Joint 2D-3D Cross-Pseudo Supervision (JCPS) method for precise carotid vessel wall segmentation.
• To address the limitations of supervised learning in 3D medical image analysis.
• To improve the accuracy and continuity of vessel wall segmentation, especially with imbalanced datasets.

Main Methods:

• A joint 2D-3D semi-supervised network was proposed to model vascular structure continuity.
• A novel four-component loss function (supervision, cross-pseudo supervision, pseudo label supervision, continuous supervision) was introduced.
• A vascular center-of-gravity positioning module was developed for automatic vessel region estimation.

Main Results:

• The JCPS method significantly outperformed the top 10 methods on the Carotid Artery Vessel Wall Segmentation Challenge dataset.
• Achieved an average Dice similarity coefficient increase from 0.775 to 0.806.
• Improved the average quantitative score from 0.837 to 0.850, demonstrating enhanced segmentation accuracy.

Conclusions:

• The JCPS method shows high generalization performance in data-imbalanced segmentation tasks.
• Pseudo labels generated by JCPS are comparable to manual software annotations.
• The proposed method offers an effective solution for accurate carotid artery vessel wall segmentation.