基于人工智能的多模态影像辅助海马体自动勾画研究

目的:利用基于深度学习的人工智能算法，结合头颅MRI和CT的多模态影像，开发海马结构自动勾画技术，为头颅放疗过程中海马体的保护提供高效、准确的自动勾画方法。方法:收集清华大学第一附属医院放疗科从2020年1月~12月就诊的40例脑转移癌患者的定位头颅CT及MRI影像，分别在CT图像、CT-MRI配准图像的两个数据集上训练3D U-Net、3D U-Net Cascade、3D BUC-Net 3个深度学习模型，计算3个模型自动分割的左右海马体与对应的人工标注之间的Dice相似系数（DSC）和95%豪斯多夫距离（95 HD），以及两者的体积作为模型的分割准确性的评估，并且以对同一大小patch图像的自动分割耗时作为模型效率的评估。结果:引入MRI图像信息对左右海马的自动分割精度有明显的提升；模型3D BUC-Net在CT-MRI数据集上对左右海马体的自动分割都取得最好分割结果（DSC:0.900±0.017，0.882±0.026；95HD:0.792±0.084，0.823±0.093），而且该模型的分割效率更高。结论:模型3D BUC-Net能在多模态影像上实现高效、准确的海马区的自动勾画，为头颅放疗过程中海马区的保护提供方便。

Auto-segmentation of the hippocampus in multimodal image using artificial intelligence

Objective To develop a technique for auto-segmentation of the hippocampal using artificial intelligence based on deep learning in the multimodal image combining magnetic resonance imaging (MRI) with computed tomography (CT), thereby providing an efficient and accurate automatic segmentation method for hippocampus sparing in cranial radiotherapy. Methods The cranial CT and MRI images of 40 patients with brain metastases treated in the Department of Radiotherapy, the First Affiliated Hospital of Tsinghua University from January 2020 to December 2020 were collected. Three kinds of deep learning models, namely 3D U-Net, 3D U-Net Cascade and 3D BUC-Net, were trained on the datasets of CT images and CT-MRI registration images separately. The Dice similarity coefficient (DSC) and 95% Hausdorff distance (95HD) between the contours of left and right hippocampus segmented automatically by models and labelde by experts, as well as the hippocampus volume were used for evaluating the segmentation accuracy of models. The time taken for auto-segmentation on the same a patch of 3D image was used to assess the efficiency of models. Results The auto-segmentation accuracy of left and right hippocampus was improved significantly by importing MRI information to CT. Among 3 kinds of models, 3D BUC-Net model had the best segmentation performance for both left and right hippocampus on CT-MRI dataset (DSC: 0.900±0.017, 0.882±0.026 95HD: 0.792±0.084, 0.823±0.093), and its segmentation efficiency was the highest. Conclusion 3D BUC-Net model can achieve more efficient and accurate automatic segmentation of the hippocampus in multimodal image, which provides a lot of convenience for the hippocampus sparing during cranial radiotherapy.