Residual-atrous attention network for lumbosacral plexus segmentation with MR image

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neuroscience

Background:

• Accurate segmentation of the lumbosacral plexus is vital for diagnosing nerve damage.
• Segmentation is challenging due to low contrast and complex anatomy in spinal MRIs.
• Existing deep learning methods often overlook nerve structures, focusing on vertebrae and discs.

Purpose of the Study:

• To develop an effective deep learning model for precise lumbosacral plexus segmentation in spinal MRIs.
• To address the limitations of current methods in segmenting low-contrast nerve structures.

Main Methods:

• Proposed a novel Residual-Atrous Attention Network (RA²-Net).
• RA²-Net incorporates an atrous encoder for multi-scale feature learning.
• Employs residual skip connections and a scale attention block for feature fusion.

Main Results:

• RA²-Net demonstrated superior performance in lumbosacral plexus segmentation.
• Achieved improved accuracy compared to existing state-of-the-art methods on a spinal MRI dataset.
• Validated on 236 MRI scans from 10 patients.

Conclusions:

• The proposed RA²-Net effectively segments the lumbosacral plexus in spinal MRIs.
• This method offers a promising advancement for clinical diagnosis and analysis of nerve damage.
• RA²-Net overcomes challenges associated with low contrast and complex anatomical structures.