A Robust Fusion Network with a Residual Multi-scale Dilated Convolution and Enhanced Attention for Muscle Identification Based on Transversus Abdominis Plane
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Ultrasound-guided regional anesthesia (UGRA) injections are widely used in peripheral nerve blocks (PNB) and have gradually become the gold standard for regional anesthesia. However, due to the high variability of ultrasound imaging and the reliance on physician experience, accurately identifying nerve structures in ultrasound images presents a significant challenge even for seasoned anesthesiologists.In this study, a multi-object assistance network for the segmentation of the transverse abdominis plane (TAP) named RMSD-Net, proposed to improve the segmentation performance in ultrasound images by simultaneously identifying the anatomical structures (e.g., muscle, bone) in the blocked region and thus improving the blocking effect. The RMSD-Net is designed based on the U-Net framework to segment and identify multiple objects simultaneously. Additionally, the Residual Multi-scale Dilated Convolution (RMDC) module is proposed to capture multi-scale feature information in ultrasound images, initially extracting the local features of the TAP. Moreover, the Skip-connection Feature Fusion (SFF) module dynamically adaptive the importance of features within each channel to enhance spatial representation capabilities. And the Multi-scale Depth-Separate Attention (MDSA) sub-module further refines the extraction of fine-grained features in the TAP by integrating spatial and channel attention mechanisms to fully exploit the incoming multi-scale features. Lastly, the Global Attention Module (GAM) is applied at the output to capture the global contextual information of the ultrasound images to prevent feature loss. The Ultrasound TAP Dataset (UTAPD) is also proposed to support the research on TAP segmentation, containing 3,000 ultrasound images with five anatomical objects (EO, IO, TA, RA, and ASIS) and their corresponding label masks. Numerous experiments on the UTAPD dataset demonstrate that RMSD-Net outperforms State-of-the-art methods in segmenting TAP muscle structures and surrounding tissues.