基于深度学习的直肠癌术后调强放疗剂量分布预测

目的 建立一种深度学习模型预测调强放疗（IMRT）的三维剂量分布。方法 收集直肠癌术后IMRT患者共110例，随机数表法选择其中90例作为训练验证集并作9折交叉验证，剩下20例作为测试集。构建3D U-Res-Net模型，以CT影像、靶区和危及器官（OARs）的解剖结构以及射束信息作为输入，IMRT剂量作为输出训练该模型，并用来预测测试集病例的剂量分布。采用三维剂量分布以及剂量—体积直方图（DVH）剂量参数评估预测精确性。结果 在三维剂量分布上，体素剂量的平均预测偏差为-2.12%~2.88%、平均绝对误差为2.55%~5.75%；等剂量面的Dice系数均在0.9以上，平均霍夫距离（HD₉₅）和平均表面距离（MSD）分别0.61~1.54 cm和0.21~0.45 cm。对于DVH剂量参数，除膀胱D_mean（P=0.048）以外，其他剂量学参数差异均无统计学意义（P>0.05）。结论 基于3D U-Res-Net模型可以实现直肠癌术后IMRT剂量分布预测，为自动计划设计奠定基础。

Dose distributions prediction for intensity-modulated radiotherapy of postoperative rectal cancer based on deep learning

Objective To develop a deep learning model for predicting three-dimensional (3D) voxel-wise dose distributions for intensity-modulated radiotherapy (IMRT). Methods A total of 110 postoperative rectal cancer cases treated by IMRT were considered in the study, of which 90 cases were randomly selected as the training-validating set and the remaining as the testing set. A 3D deep learning model named 3D U-Res-Net was constructed to predict 3D dose distributions. Three types of 3D matrices from CT images, structure sets and beam configurations were fed into the independent input channel, respectively, and the 3D matrix of IMRT dose distributions was taken as the output to train the 3D model. The obtained 3D model was used to predict new 3D dose distributions. The predicted accuracy was evaluated in two aspects:the average dose prediction bias and mean absolute errors (MAEs)of all voxels within the body, the dice similarity coefficients (DSCs), Hausdorff distance(HD₉₅) and mean surface distance (MSD) of different isodose surfaces were used to address the spatial correspondence between predicted and clinical delivered 3D dose distributions; the dosimetric index (DI) including homogeneity index, conformity index,V₅₀,V₄₅ for PTV and OARs between predicted and clinical truth were statistically analyzed with the paired-samples t test. Results For the 20 testing cases, the average prediction bias ranged from -2.12% to 2.88%, and the MAEs varied from 2.55% to 5.75%. The DSCs value was above 0.9 for all isodose surfaces, the average MSD ranged from 0.21 cm to 0.45 cm, and the average HD₉₅ varied from 0.61 cm to 1.54 cm. There was no statistically significant difference for all DIs, except for bladder D_mean. Conclusions This study developed a deep learning model based on 3D U-Res-Net by considering beam configurations input and achieved an accurate 3D voxel-wise dose prediction for rectal cancer treated by IMRT.