基于深度学习方法的食管癌术后调强放疗三维剂量分布预测

目的:构建一种深度学习网络模型预测食管癌调强放疗的三维剂量分布。方法:取100例中上段食管癌术后患者的调强放疗计划为研究对象，以患者计划的计算机断层扫描（CT）图像、靶区和危及器官的勾画图像以及适形射束信息作为输入数据，调强适形放射治疗（IMRT）的三维剂量分布作为输出数据，通过搭建的3D U-Res-Net混合网络进行训练并得到预测模型，利用该模型对测试集进行三维剂量预测。采用平均预测偏差δ]、平均绝对误差（MAE）、戴斯相似性系数（DSC）和豪斯多夫距离（HD95）评价预测结果的精确性。结果:测试集的平均预测偏差为-0.23%~0.78%，MAE为1.67%~3.07%，两组计划等剂量面DSC均值大于0.91，尤其30 Gy以下的DSC达到0.95以上，平均HD95为0.51~0.73 cm。预测计划的剂量学参数均在临床允许的范围之内且相对剂量偏差小于2%，除靶区D2、脊髓Dmax、全肺V30差异有统计意义外（P<0.05），其余剂量学参数差别不大。结论:本研究构建的3D U-Res-Net深度学习网络模型可以实现对食管癌术后IMRT三维剂量分布的精确预测。

Deep learning-based prediction of three-dimensional dose distribution in postoperative intensity-modulated radiotherapy for esophageal cancer

Objective To develop a deep learning network model for predicting the three-dimensional(3D)dose distribution in postoperative intensity-modulated radiotherapy(IMRT)for esophageal cancer.Methods A total of 100 postoperative patients with upper and middle esophageal cancer treated by IMRT were enrolled in the study.The CT images,segmentations of target areas and organs-at-risk,and conformal beam configuration were taken as input data,and IMRT dose distribution was taken as output data.The established hybrid network 3D U-Res-Net was used for training and obtaining prediction model which was then used for the prediction of 3D dose distribution on the test set.The prediction accuracy was evaluated by the average prediction bias-δ,mean absolute error(MAE),Dice similarity coefficient(DSC)and Husdorff distance(HD₉₅).Results For the test set,the average prediction bias ranged from-0.23%to 0.78%,and MAE varied from 1.67%to 3.07%.The average DSC was above 0.91 for all isodose surfaces,especially when the dose was less than 30 Gy(DSC was higher than 0.95),and the average HD95was from 0.51 cm to 0.73 cm.The dosimetric parameters of the prediction plan were all within the clinically allowable range,and the relative dose deviation was less than 2%.There is no significant difference in dosimetric parameters except for D₂ to target area,D_max to spinal cord and V₃₀ of whole lung(P<0.05).Conclusion The 3D dose distribution in the postoperative intensity-modulated radiotherapy(IMRT)for esophageal cancer can be accurately predicted by the established 3D U-Res-Net model.