CervSpineNet: a hybrid deep learning-based approach for the segmentation of cervical spinous processes

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Accurate segmentation of cervical spinous processes is crucial for spinal analysis but lacks public datasets and efficient methods.
• Manual delineation is time-consuming and operator-dependent, hindering clinical and research applications.

Purpose of the Study:

• To develop an automated, accurate, and efficient method for segmenting cervical spinous processes on lateral X-rays.
• To create and release an expert-annotated dataset for cervical spine analysis.

Main Methods:

• Developed CervSpineNet, a hybrid deep learning framework integrating transformer and convolutional components.
• Trained the model using a compound loss function optimizing multiple segmentation metrics (Dice, Focal Tversky, HDT, SSIM).
• Evaluated performance against U-Net, DeepLabV3+, SAM, and SegFormer on diverse image sets.

Main Results:

• CervSpineNet achieved superior performance, with Dice coefficients >0.93, IoU >0.87, and SSIM >0.98.
• The model demonstrated high accuracy (MAE ≈ 0.005) and efficient inference (5-10 seconds per image).
• Reduced manual annotation time by approximately 96% with a compact model size.

Conclusions:

• The combination of global context (transformer) and local refinement (convolutional) enables robust spinous process segmentation.
• CervSpineNet offers a scalable foundation for AI-assisted cervical spine analysis in clinical and research settings.
• This work provides a high-performing, efficient model and an expert-annotated dataset, accelerating AI applications in spinal imaging.