基于U-net卷积神经网络的宫颈癌临床靶区和危及器官自动勾画的研究

Automatic segmentation of clinical target volumes and organs-at-risk in radiotherapy for cervical cancer using U-net convolutional neural network

目的:基于U-net卷积神经网络的深度学习方法,探讨宫颈癌放疗临床靶区和危及器官自动勾画的可行性。方法:利用U-net卷积神经网络模型搭建的端到端自动分割框架,以100例已进行IMRT治疗的宫颈癌患者CT及组织结构信息为研究对象,并随机选取其中的10例作为测试集。勾画的对象包括临床靶区(CTV)、膀胱、直肠和左、右股骨头5个部分,比较手动和自动勾画的戴斯相似性系数(DSC)和豪斯多夫距离(HD)以评估自动勾画模型的准确性。结果:4种危及器官自动勾画的DSC值都在0.833以上,平均值是0.898;HD值均在8.3 mm以内,平均值为5.3 mm;临床靶区DSC值是0.860,HD值为13.9 mm。结论:基于U-net卷积神经网络建立的自动勾画模型能较为准确地实现宫颈癌临床靶区和危及器官的自动勾画,临床应用中可大幅提高医生的工作效率及勾画的一致性。

Objective To explore the feasibility of deep learning based on U-net convolutional neural network for the automatic segmentation of clinical target volumes and organs-at-risk in the radiotherapy for cervical cancer. Methods U-net convolutional neural network model was used to construct an end-to-end automatic segmentation framework. The CT and tissue structure data of 100 patients with cervical cancer who had undergone intensity-modulated radiotherapy were analyzed in this study, and 10 of the patients were randomly selected as test sets. The clinical target volume, the bladder, the rectum and the left and right femoral heads were segmented. Dice similarity coefficient(DSC) and Hausdorff distance(HD) of manual and automatic segmentations were compared to evaluate the accuracy of the automatic segmentation model. Results All the DSC of organs-at-risk was above 0.833, with an average value of 0.898;and all the HD was within 8.3 mm, with an average value of 5.3 mm. The DSC and HD of clinical target volumes were 0.860 and 13.9 mm, respectively. Conclusion The automatic segmentation model established based on U-net convolutional neural network can accurately realize the automatic segmentations of clinical target volumes and organs-at-risk in the radiotherapy for cervical cancer, and it can also greatly improve the working efficiency of doctors and the consistency of segmentations in clinical application.