DEMAC-Net: A Dual-Encoder Multiattention Collaborative Network for Cervical Nerve Pathway and Adjacent Anatomical...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Cervical anesthesia techniques carry risks and require expertise.
• Ultrasound imaging is valuable but struggles with segmenting small neural structures.
• Accurate segmentation is crucial for safe and effective ultrasound-guided anesthesia.

Purpose of the Study:

• Introduce DEMAC-Net, a dual-encoder, multi-attention network for improved segmentation of cervical and brachial plexuses.
• Enhance the identification of cervical nerve pathways (CNP) and adjacent tissues.
• Aid clinicians in guiding anesthesia procedures and optimizing needle insertion points.

Main Methods:

• Utilized a dual-encoder architecture with Spatial Understanding Convolution Kernel (SUCK) and Spatial-Channel Attention Module (SCAM).
• Incorporated Global Attention Gate (GAG) and inter-layer fusion for feature refinement and noise suppression.
• Developed a new Neck Ultrasound Dataset (NUSD) with 1,500 annotated images and validated on the BUSI dataset.

Main Results:

• DEMAC-Net achieved a 93.3% Dice Similarity Coefficient (DSC) on the NUSD dataset.
• Demonstrated superior generalization with 87.2% DSC and 77.4% Intersection over Union (IoU) on the BUSI dataset.
• Showcased consistent segmentation stability across various anatomical structures.

Conclusions:

• DEMAC-Net significantly enhances segmentation accuracy for small nerves and complex structures in ultrasound images.
• The network outperforms existing methods in accuracy and computational efficiency.
• This framework has the potential to improve ultrasound-guided procedures like peripheral nerve blocks, leading to better clinical outcomes.