Research on the 3D Dose Prediction Based on Deep Learning for Rectal Cancer
Volumetric Modulated Arc Therapy
LIU Runhong1, LIU Ke2, HUANG Qiang1, XU Xiaoming1, XU Hui3
1. Department of Radiotherapy, Neijiang Second People’s Hospital, Neijiang Sichuan 641000, China; 2. Department of Oncology,
Zigong First People’s Hospital, Zigong Sichuan 643000, China; 3. Department of Radiology, Neijiang Sixth People’s Hospital,
Neijiang Sichuan 641000, China
Abstract：Objective To propose a 3DRes-UNet deep learning network for predicting the 3D dose accuracy of postoperative volume
modulated arc therapy (VMAT) for rectal cancer surgery, so as to guide clinical practice. Methods A total of 168 VMAT radiotherapy
plans for rectal cancer was collected. The data set was randomly divided into a training set of 120 cases, a validation set of 16
cases, and a test set of 32 cases in a 7∶1∶2 ratio. The CT images of the training set and masks of organs and target volume were
input into the network for training. The predicted dose was compared with clinically approved radiotherapy doses on the test set to
evaluate the accuracy of radiotherapy dose prediction. Results There was no statistically significant difference in D2, D98, D50, and
homogeneity index between the clinical dose and predicted values in the target volume (P>0.05). There was a statistical difference
in the conformity index (P<0.05). The predicted doses of V50 and Dmean for organ threatening bladder were lower than clinical doses
(P<0.05), and there was no statistically significant difference in V40 (P>0.05). The predicted dose of V40 in the left femoral head
was lower than the clinical dose (P<0.05), and there was no statistically significant difference in V30, V50, and Dmean (P>0.05). The
predicted dose of Dmean in the right femoral head was lower than the clinical dose (P<0.05), and there was no statistically significant
difference in V30, V40, and V50 (P>0.05). The predicted doses of pelvic V45 and Dmean were also lower than clinical doses (P<0.05).
There was no statistically significant difference in V30, V40, Dmean, and D0.1cc in the small bowel (P>0.05). The dose difference map
showed that there was little difference between the predicted results of the target area and the clinical results, and the difference in
organ endangerment was between -10-10 Gy. The predicted dose volume histogram basically coincided with the clinical dose volume
histogram. Conclusion The 3DRes-UNet model can effectively predict the 3D space dose of postoperative VMAT radiotherapy plan
for rectal cancer, and guide clinical radiotherapy work.
刘润红1，刘可2，黄强1，徐孝明1，许惠3. 基于深度学习的直肠癌术后容积旋转调强
放疗三维剂量预测研究[J]. 中国医疗设备, 2024, 39(4): 41-46.
LIU Runhong1, LIU Ke2, HUANG Qiang1, XU Xiaoming1, XU Hui3. Research on the 3D Dose Prediction Based on Deep Learning for Rectal Cancer
Volumetric Modulated Arc Therapy. China Medical Devices, 2024, 39(4): 41-46.
