面向胸部骨分割的混合双编码器模型及应用

Hybrid Dual Encoder Model for Thoracic Bone Segmentation and Its Application

针对现有的胸部骨分割模型无法达到肋骨配准对分割精度的要求和数据集无法满足肋骨配准研究的问题.本文首先构建了一个面向肋骨刚性配准的胸部骨分割数据集,同时对现有的分割模型进行了改进,最后搭建了一套完整的肋骨配准框架.首先,三维胸部CT图像数据以MICCAI 2020 RibFrac挑战赛上的公开数据为基础,在胸科医生指导下,采用半自动的方式对4类骨头进行标注.然后,针对卷积神经网络无法对图像的长距离语义关系建模的问题,在VB-Net的基础上,实现了一种CNN与Transformer相结合的多类胸部骨自动分割网络.最后,为提升肋骨运动建模的效率,以肋骨分割为基础,设计了肋骨刚性配准算法,实现同一个患者的两期图像的肋骨自动对齐.实验结果表明,与现有的分割算法相比,改进的算法在分割Dice、分割结果可视化方面更优,此外,所提出的配准框架能够高效实现肋骨的自动对齐.

Aiming at the problem that the existing chest bone segmentation model can not meet the requirement of segmentation accuracy and the dataset can not meet the research of rib registration.In this paper,we construct a chest bone segmentation data set for rib rigid registration,improve the existing segmentation model,and finally build a complete frame for rib registration.First,3D chest CT image data was based on open data from the MICCAI 2020 RibFrac Challenge.Under the guidance of the chest physician,four types of bones were labeled in a semi-automatic manner.Then,in view of the problem that convolutional neural network cannot model long-distance semantic relationship of image,a multi-class bone automatic segmentation network combined with CNN and Transformer is implemented on the basis of VB-Net.Finally,in order to improve the efficiency of rib motion modeling,a rigid rib registration algorithm was designed based on rib segmentation to realize automatic rib alignment between the CT images of two breath phases belonging to the same patient.Experimental results show that compared with existing segmentation algorithms,the improved algorithm achieves the better performance in terms of Dice and visualization of segmentation results.In addition,the proposed registration framework can realize automatic alignment of ribs with high quality.