基于级联3D U-Net的CT和MR视交叉自动分割方法

目的:基于级联3D U-Net，利用配对患者头颈部数据［CT和磁共振图像（MRI）］，取得比仅CT数据更高分割精度的视交叉自动分割结果。方法:该级联3D U-Net由一个原始3D U-Net和改进的3D D-S U-Net（3D Deeply-Supervised U-Net）组成，实验使用了60例患者头颈部CT图像及MRI图像（T1和T2模态），其中随机选取15例患者数据作为测试集，并使用相似性系数（DSC）评估视交叉的自动分割精度。结果:对于测试集中的所有病例，采用多模态数据（CT和MRI）的视交叉的DSC为0.645±0.085，采用单模态数据（CT）的视交叉的DSC为0.552±0.096。结论:基于级联3D U-Net的多模态自动分割模型能够较为准确地实现视交叉的自动分割，且优于仅利用单模态数据的方法，可以辅助医生提高放疗计划制定的工作效率。

Automatic optic chiasm segmentation using CT and MRI based on cascaded 3D U-Net

Abstract: Objective To realize the automatic segmentation of the optic chiasm using multimodal images (CT and MRI) that contain head-and-neck data and based on cascaded 3D U-Net for obtaining a higher segmentation accuracy than using only CT data. Methods The proposed cascaded 3D U-Net consists of an original 3D U-Net and an improved 3D D-S U-Net (3D Deeply-Supervised U-Net). The head-and-neck CT images and MRI images (T1 and T2 modalities) of 60 patients were used in the experiment, and the data of 15 patient were randomly selected as the test set. Dice similarity coefficient was used to evaluate the accuracy of automatic optic chiasm segmentation. Results For all cases in the test set, the Dice similarity coefficient of the optic chiasm segmentation using multimodal data (CT and MRI) or monomodal data (CT) was 0.645±0.085 and 0.552 ±0.096, respectively. Conclusion The multimodal automatic segmentation model based on cascaded 3D U-Net can accurately realize the automatic segmentation of the optic chiasm, superior to the method using only monomodal data, and it can assist doctors in improving the efficiency of radiotherapy planning.