VerFormer: Vertebrae-Aware Transformer for Automatic Spine Segmentation from CT Images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate spine segmentation is crucial for diagnosing spinal conditions but remains challenging due to anatomical variations.
• Existing methods using Convolutional Neural Networks (CNNs) struggle to capture global contextual information, limiting segmentation accuracy.
• Vision Transformers (ViTs) can capture global context but treat all image regions equally, lacking focus on relevant spinal structures.

Purpose of the Study:

• To develop an improved automatic spine segmentation method for CT images.
• To enhance the utilization of global contextual information for more accurate spine segmentation.
• To address the limitations of existing CNN and ViT models in capturing vertebrae-specific features.

Main Methods:

• Proposed a novel Vertebrae-aware Vision Transformer (VerFormer) model for spine segmentation.
• Introduced a Vertebrae-aware Global (VG) block with a Vertebrae-aware Global Query (VGQ) module.
• Integrated VG block into the Vision Transformer backbone to highlight vertebrae-related tokens using multi-head self-attention.

Main Results:

• The VerFormer model demonstrated superior capacity in capturing discriminative dependencies and vertebrae-related context.
• Experimental results on two spine CT segmentation tasks confirmed the effectiveness of the VG block.
• VerFormer achieved higher segmentation accuracy and better generalization compared to popular CNN- and ViT-based models.

Conclusions:

• The proposed Vertebrae-aware Vision Transformer (VerFormer) significantly advances automatic spine segmentation from CT images.
• The Vertebrae-aware Global (VG) block effectively leverages global contextual information for improved localization and segmentation.
• VerFormer offers a superior alternative for clinical applications requiring precise spine segmentation.