基于深度学习方法的乳腺癌调强放疗自动计划研究

Study of automatic treatment planning of intensity-modulated radiotherapy based on deep learning technique for breast cancer patients

目的开发一种基于深度学习网络的乳腺癌调强放疗计划剂量分布预测的方法,并评估将其用于自动计划的可行性。方法从复旦大学附属肿瘤医院选取240例左侧乳腺癌患者,200例作为训练集,20例作为验证集,另外20例作为测试集。应用深度学习网络建立患者CT影像、靶区和危及器官的勾画图像与剂量分布的相互关系,达到预测新患者剂量分布的目的,并尝试将预测的剂量分布作为目标函数优化并生成治疗计划。结果临床治疗计划的剂量分布和预测的剂量分布相比,靶区(除同步加量的PTV48Gy)和危及器官的剂量值相近,且基于预测的剂量分布生成的治疗计划与预测结果基本相同。结论本研究实现了一种基于深度学习网络的乳腺癌调强计划剂量分布预测方法,有助于进一步实现自动设计治疗计划的目标。

Objective To develop a deep learning-based approach for predicting the dose distribution of intensity-modulated radiotherapy(IMRT)for breast cancer patients,and evaluate the feasibility of applying the predicted dose distribution in the automatic treatment planning.Methods A total of 240 patients with left breast cancer admitted to Fudan University Shanghai Cancer Center were enrolled in this study:200 cases in the training dataset,20 cases in the validation dataset and 20 cases in the testing dataset.A modified deep residual neural network was trained to establish the relationship between CT image,the contouring images of target area and organs at risk(OARs)and the dose distribution,aiming to predict the dose distribution.The predicted dose distribution was utilized as the optimization objective function to optimize and generate a high-quality plan.Results Compared with the dose distribution of clinical treatment plan,the predicted dose distribution for target areas and OARs showed no statistical significance except for a simultaneous boost target PTV48Gy.And the treatment plan generated based on the predicted dose distribution was basically consistent with the predicted outcomes.Conclusion Our results demonstrate that the deep learning-based approach for predicting the dose distribution of IMRT for breast cancer contributes to further achieving the goal of automatic treatment planning.