构建混合多尺度神经网络自动分割Graves眼病的放疗临床靶区

目的：构建混合多尺度神经网络（HMnet）实现放疗临床靶区的自动勾画，提供一个高精度的CT影像自动分割模型。方法：HMnet是一种端到端的卷积神经网络，使用深度残差网络提取特征，由4个不同内核的卷积层组成的多尺度特征融合模块进行处理，以适应不同尺度大小的临床靶区；再引入注意力残差模块对多尺度特征融合模块输出的有效特征进行强化。共采用117例Graves眼病病例的CT影像数据和临床靶区轮廓训练和评估HMnet，选择骰子相似系数（DSC）、95%豪斯多夫距离（95HD）作为评估指标。结果：采用HMnet进行Graves眼病放疗临床靶区自动勾画的DSC为0.874 9,95HD为2.525 4 mm，均优于Unet、Vnet、ResAttUnet3D网络，也优于同一位医生两次勾画结果的平均DSC。结论：HMnet能准确实现Graves眼病放疗临床靶区的自动勾画，可提高放疗医生的工作效率及勾画的一致性。

Construction of a hybrid multi-scale neural network for auto-segmentation of the clinical target areas in Graves ophthalmopathy

Abstract: Objective To construct a hybrid multi-scale neural network (HMnet) to automatically delineate the clinical target volumes (CTV) for providing a high-precision model for the auto-segmentation in CT images. Methods HMnet which was an end-to-end convolutional neural network used the deep residual network to extract features, carried out feature processing with multi-scale feature fusion module composed of 4 convolution layers of different cores to adapt to the clinical target volumes of different scales, and introduced the attention residual module to enhance the effective features output by the multi-scale feature fusion module. HMnet was trained and tested using the CT data and CTV contours from 117 cases of Graves ophthalmopathy. The Dice similarity coefficient (DSC) and 95% Hausdorff distance (95HD) were selected as evaluation metrics. Results HMnet (DSC=0.874 9, 95HD=2.525 4 mm) outperformed Unet, Vnet and ResAttUnet3D, and the DSC of HMnet was also higher than the average DSC of two delineation results from the same oncologists. Conclusion HMnet can accurately achieve the automated CTV segmentation in Graves ophthalmopathy, and its clinical application can improve the efficiency of oncologists and the consistency of delineation.