基于多尺度边缘分割与混合注意力机制的脊柱CT图像分割

Spine CT image segmentation based on multi-scale boundary segmentation and hybrid attention mechanism

脊柱疾病的前期主要通过计算机断层扫描技术进行筛查与初步判断。为解决脊柱CT图像目前存在的椎骨结构复杂、分割精度不足等问题,提出一种基于3D U-Net框架的脊柱CT图像改进分割网络,通过融合SE残差单元、椎骨边缘分割模型与改进混合通道-空间注意力机制,在VerSe 19、VerSe 20与CTSpine1K脊柱数据集上进行分割训练与测试。多次测试实验结果表明,本文模型在保证分割精度和分割效率有效提高的同时具有较好的泛化性与鲁棒性,在Dice相似系数、豪斯多夫距离与平均表面距离上相较于其他先进网络分割精度更高。本文模型在现有脊柱分割的网络中具有更强的分割性能,可为放射科医生提供有效临床信息。

The early diagnosis of spinal diseases is mainly screened and initially diagnosed through computed tomography(CT).In view of the complex structure of vertebral bones and low segmentation accuracy in spinal CT images,a spinal CT image segmentation network based on 3D U-Net framework is proposed.The network which integrates squeeze-andexcitation residual module,vertebral boundary segmentation model,and improved hybrid channel-spatial attention mechanism is trained and tested on VerSe 19,VerSe 20,and CTSpine1K spinal datasets.Multiple experiments indicate that the model can effectively improve segmentation accuracy and efficiency while demonstrating good generalization and robustness.Compared with other advanced network models,the proposed network achieves higher segmentation accuracy in terms of Dice similarity coefficient,Hausdorff distance,and average symmetric surface distance.The proposed model exhibits superior segmentation performance among the existing spinal segmentation networks,providing radiologists with valuable clinical information.