Deep Learning-Based Automatic Segmentation of Lumbosacral Nerves on CT for Spinal Intervention: A Translational Study

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• 3D reconstruction of surgical "safe" and Kambin triangles aids epidural steroid injection viability assessment, particularly at L5/S1.
• Manual segmentation of lumbosacral nerves for 3D models is labor-intensive and time-consuming.

Purpose of the Study:

• To assess the feasibility of deep learning for segmenting lumbosacral nerves on CT.
• To validate 3D reconstruction of the safe and Kambin triangles using deep learning segmentation.

Main Methods:

• 50 spinal CT cases were manually labeled for nerves and bones.
• A 3D U-Net model (SPINECT) was developed for automated segmentation.
• Segmentation performance was evaluated using Dice score, pixel accuracy, and Intersection over Union.

Main Results:

• SPINECT achieved high accuracy in segmenting lumbosacral bone (Dice: 0.945, IoU: 0.897) and nerves (Dice: 0.905, IoU: 0.827).
• Quantified areas of Kambin and safe triangles showed no significant difference between manual and automated segmentation (P > .05).

Conclusions:

• Deep learning-based automatic segmentation of lumbosacral structures on CT is feasible.
• SPINECT facilitates validated 3D reconstruction of surgical triangles for improved anatomical assessment.