Benefits of deep learning for delineation of organs at risk in head and neck cancer
Benefits of deep learning for delineation of organs at risk in head and neck cancer
复制标题
DOI：
10.1016/j.radonc.2019.05.010
复制
发表时间：
2019-09-01
影响因子：
5.7
通讯作者：
Nuyts, S.
中科院分区：
文献类型：
--
作者：
van der Veen, J.;Willems, S.;Nuyts, S.
目的：精确勾画头颈部癌(HNC)的危险器官(OARs)是精确放射治疗的必要条件。虽然有指导方针，但观察者间的显著变异性(IOV)仍然存在。目的是验证三维卷积神经网络(CNN)半自动勾画桨叶的准确性、效率和一致性，并与手工勾画进行比较。材料/方法：由两名训练有素的放射肿瘤学家(RO)按照国际公认的指南，手动勾画15名新的HNC患者的16个桨叶。桨也通过应用CNN自动划定，并分别根据两个RO的需要进行修正。这两种描述都是在两周内进行的，并且彼此都是盲目的。用Dice相似系数(DSC)和平均对称表面距离(ASSD)来量化两种RO之间的IOV。结果：自动勾画的平均校正时间比人工勾画的平均时间缩短33%(23比34分钟)(p<；10-6)。随着几乎所有桨的网络初始化，IOV显著改善(p<；0.05)，导致所有桨的平均ASD从1.9毫米下降到1.2毫米。该网络对所有桨的RO1和RO2的平均DSC准确率分别为90%和84%，ASSD分别为0.7和1.5 mm，分别低于IOV的93%和73%。结论：为HNC自动绘制OAR而开发的CNN比手动绘制更加有效和一致，在临床上是可行的。(C)2019爱思唯尔B.V.保留所有权利。
Purpose/objective: Precise delineation of organs at risk (OARs) in head and neck cancer (HNC) is necessary for accurate radiotherapy. Although guidelines exist, significant interobserver variability (IOV) remains. The aim was to validate a 3D convolutional neural network (CNN) for semi-automated delineation of OARs with respect to delineation accuracy, efficiency and consistency compared to manual delineation.Material/methods: 16 OARs were manually delineated in 15 new HNC patients by two trained radiation oncologists (RO) independently, using international consensus guidelines. OARs were also automatically delineated by applying the CNN and corrected as needed by both ROs separately. Both delineations were performed two weeks apart and blinded to each other. IOV between both ROs was quantified using Dice similarity coefficient (DSC) and average symmetric surface distance (ASSD). To objectify network accuracy, differences between automated and corrected delineations were calculated using the same similarity measures.Results: Average correction time of the automated delineation was 33% shorter than manual delineation (23 vs 34 minutes) (p < 10-6). IOV improved significantly with network initialisation for nearly all OARs (p < 0.05), resulting in decreased ASSD averaged over all OARs from 1.9 to 1.2 mm. The network achieved an accuracy of 90% and 84% DSC averaged over all OARs for RO1 and RO2 respectively, with an ASSD of 0.7 and 1.5 mm, which was in 93% and 73% of the cases lower than the IOV.Conclusion: The CNN developed for automated OAR delineation in HNC was shown to be more efficient and consistent compared to manual delineation, which justify its implementation in clinical practice. (C) 2019 Elsevier B.V. All rights reserved.