CT-UNet: A Convolutional-Transformer Network for Enhanced Edge Detection in Artificial Vertebrae Segmentation

Spinal metastases often require surgical resection with 3D-printed artificial vertebrae for reconstruction. Precise segmentation is crucial for assessing bone fusion via CT Hounsfield Unit (HU) analysis, yet existing methods struggle with irregular implant shapes and edge details. While convolutional neural networks (CNNs) and Transformers dominate spinal segmentation, they have distinct limitations: CNNs lack global context, and Transformers struggle with local details and high computational costs. Existing studies primarily focus on accurately locating vertebrae in images and determining their relative positions to adjacent vertebrae, with insufficient attention given to vertebral edge details and no public datasets address this gap. In this paper, we construct a spine CT image dataset containing artificial vertebrae (AV SpineCT) and propose a novel segmentation model, CT-UNet. CT-UNet is an end-to-end U-Net model that integrates a Convolutional-Transformer Encoder (CT Encoder) and a Spatial-Channel attention module based on edge prior knowledge (ESCA). Additionally, we introduce Edge-based Focal Loss, designed to enhance focus on edge information and mitigate blurred edge segmentation issues. Experimental results demonstrate that CT-UNet outperforms existing models, with a 5.61% improvement in edge segmentation accuracy (Boundary IoU) compared to TransUnet.