SCAR-Net: Spine Segmentation in MRI based on Cross Attention and Recognition-assisted Label Fusion

The segmentation of multiple vertebrae and intervertebral discs in magnetic resonance images (MRI) plays a crucial role in diagnosing and treating spinal disorders. However, the inherent complexity of the spine, coupled with the challenges of balancing inter-class similarity and intra-class variety, complicates the task. Additionally, improving the generalization ability, learning rate, and accuracy of spine segmentation remains difficult. To address these challenges, this paper proposes a spine segmentation method based on cross attention and recognition-assisted label fusion (SCAR-Net). The approach introduces a multi-channel cross attention (MCCA) mechanism to generate a comprehensive spine description by fusing inter-class and intra-class features. Furthermore, a key-points recognition-assisted learner (KRAL) is designed, incorporating mixed-supervision recognition-assisted label fusion (RALF) to reduce reliance on a single dataset and enhance network generalization. Experimental results on T2-weighted volumetric MRI datasets demonstrate that SCAR-Net achieves outstanding performance, with a mean Dice similarity coefficient (DSC) of 96.12% for 5 vertebral bodies and 95.07% for 5 intervertebral discs. The proposed method proves to be highly effective for both the localization and segmentation of intervertebral discs in MRI spine images.