Towards accurate facial nerve segmentation with decoupling optimization.

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Robotic cochlear implantation restores hearing but relies on accurate facial nerve identification.
• Facial nerve segmentation is challenging due to small size, similar surrounding areas, and low contrast.

Purpose of the Study:

• To develop an effective end-to-end neural network for precise facial nerve segmentation.
• To address the challenges of small object detection and low contrast in medical images.

Main Methods:

• Proposed FNSegNet, a two-stage neural network for facial nerve segmentation.
• Coarse segmentation stage: utilized search identification modules and pyramid fusion.
• Refine segmentation stage: employed decoupling optimization and spatial attention modules.

Main Results:

• Achieved significant improvements in segmentation accuracy (0.858 Dice, 0.363 mm 95% Hausdorff distance).
• Reduced computational complexity (13.33 GFLOPs, 9.86 M parameters).
• Demonstrated effectiveness on a challenging dataset.

Conclusions:

• FNSegNet offers a robust solution for accurate facial nerve segmentation in robotic surgery.
• The proposed method enhances the safety and efficacy of cochlear implantation procedures.
• This AI-driven approach advances precision in neurosurgical interventions.