DMRNet: a dynamic multi-scale residual network for Shamrock view and lumbar plexus segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anesthesiology

Background:

• Lumbar plexus block (LPB) is crucial for hip and knee surgeries.
• Ultrasound guidance is used, but lumbar plexus anatomy presents challenges.
• Accurate sonoanatomical recognition is vital for effective LPB.

Purpose of the Study:

• To develop a deep learning model for precise lumbar plexus segmentation in ultrasound images.
• To assist anesthesiologists in identifying the lumbar plexus in the Shamrock view.
• To improve the accuracy and efficiency of ultrasound-guided LPB.

Main Methods:

• Proposed DMRNet, a deep learning model integrating Adaptive Multi-Scale Dilated (AMD) and Dense Attention Residual (DAR) modules.
• Incorporated Attention-Enhanced Hybrid (AEH) Module and attention mechanisms (BASA, ER-MHA).
• Trained and evaluated the model on Shamrock view ultrasound images for delineating muscles, nerves, and bony structures.

Main Results:

• DMRNet achieved a mean Intersection over Union (IoU) of 0.863.
• DMRNet achieved a mean Dice coefficient of 0.926 across all target structures.
• The model outperformed existing state-of-the-art segmentation models.

Conclusions:

• DMRNet demonstrates high accuracy in segmenting sonoanatomical structures in the Shamrock view.
• The model shows potential as an assistive tool for anesthesiologists during ultrasound-guided LPB.
• DMRNet can offer educational support for training in ultrasound-guided LPB.