基于U-Net结合改进算法对放疗危及器官自动勾画研究

目的：面向放疗危及器官自动勾画构建基于U-Net的模型并针对肝脏分割构建3种改进模型。方法：采集共计184例肝癌患者和183例头部放疗患者的计算机断层扫描（CT）图像及组织结构信息，并结合公开数据集Sliver07用于模型的训练与评估。通过搭建U-Net模型并针对肝脏分割分别结合空洞卷积、SLIC超像素算法、区域生长算法进行训练并得到预测模型，利用预测模型对自动勾画结果进行预测。采用交并比（Io U）和平均交并比（MIo U）评价预测结果的精确性。结果：测试集头部放疗危及器官自动勾画预测结果MIo U为0.795～0.970，肝脏分割使用U-Net预测结果MIo U约为0.876，使用改进后模型预测结果MIo U约为0.888，并很好地约束了预测偏差较大结果的出现，使得测试样本中Io U结果小于0.8的数量占比从16.67%降至7.5%。直观勾画方面结合改进算法的模型比U-Net更能捕捉到复杂、混淆性的边界区域。结论：构建U-Net模型能够在头部放疗危及器官和肝脏自动勾画上表现良好，3种改进的模型能够在肝脏分割上具有更优的表现。

Auto-segmentation of organs-at-risk for radiotherapy using U-Net combined with improved algorithms

Abstract: Objective To develop a model based on U-Net for the auto-segmentation of organs-at-risk, and to propose 3 improved models for automated liver segmentation. Methods The CT images and tissue structure data of 184 patients with liver cancer and 183 patients receiving head radiotherapy were collected and combined with the public dataset Sliver07 for the training and evaluation of the models. The established U-Net model and 3 models combined with dilated convolution, SLIC super-pixel algorithm and region growing algorithm, respectively, were trained for obtaining prediction models which were then used for the prediction of auto-segmentation results. The segmentation accuracy was evaluated using intersection over union (IoU) and mean intersection over union (MIoU). Results For the test set, the MIoU of the U-Net model for OAR segmentation in head radiotherapy ranged from 0.795 to 0.970 and for the liver segmentation was around 0.876. The improved model for automated liver segmentation improved the MIoU to 0.888 and restricted the occurrence of large prediction deviations, which reduced the proportion of IoU less than 0.8 in the test samples from 16.67% to 7.50%. Visually, the models combined with improved algorithms could capture more complex and confusing boundary areas than U-Net. Conclusion The established U-Net performed well in the auto-segmentations of liver and organs-at-risk for head radiotherapy, and the 3 improved models can obtain better results in liver segmentation.