X

历史悠久的传统与大数据和机器学习时代的碰撞

编译:肿瘤资讯

来源:肿瘤资讯

头颈部放疗靶区勾画:历史悠久的传统与大数据和机器学习代的碰撞

Melvin Lee Kiang Chua教授

新加坡国立癌症中心(NCCS)放疗科 临床科学家、高级顾问医师

精准放疗计划项目(NCCS)负责人

本次报告,Melvin教授主要从传统靶区概念、国际指南,以及目前迅速发展的机器学习和人工智能(AI)角度阐述了如何实现加精准的放疗靶区勾画,从而保证患者的生存。他指出头颈部精准放疗需要稳定的质量保证,而精确的靶区勾画是其中的重要环节。以鼻咽癌为例,目前传统的原发肿瘤大体肿瘤靶区(GTV)、临床靶区(CTV)勾画中存在一定的问题。

对于原发肿瘤GTV而言,影像手段(CT、MRI、PET-CT)的进步使得勾画准确性提高。然而,医生对解剖结构“异常”和“正常”的认定相对主观;因此,即使在经验丰富的放疗医生之间,GTV勾画仍然存在高度异质性,准确性高度依赖医生的经验;并且勾画过程十分耗时耗力。利用深度卷积神经网络技术实现GTV自动勾画将有助于改善这些问题。中山大学肿瘤防治中心的研究显示利用三维卷积神经网络在多参数MR图像上进行GTV自动勾画,其准确性超过参与研究的半数医生(5/8)。在AI自动勾画辅助下,勾画者间差异减少一半,勾画时间缩短40%。

对于颈部CTV勾画而言,能否减少预防照射区域或适当降低剂量是目前研究的热点。最近,依靠大数据的力量,林丽等人在模板CT上标记了959例患者的10651颗淋巴结,建立了鼻咽癌淋巴结分布图谱并对2013版头颈肿瘤淋巴引流区勾画的国际指南指南提出了7点针对鼻咽癌的修改建议。

对于原发灶CTV而言,2017年发表的国际专家共识指南能够为减少医生之间的不确定性提供指导。然而,Melvin教授认为,在大数据和机器学习的时代,我们应该考虑能否利用大数据优化对CTV范围的界定呢?因此,探索使用机器学习的方法,利用大量原发肿瘤侵犯范围的数据,通过计算未受侵犯的CT体素可能侵犯的概率从而生成基于概率的CTV将有可能进一步帮助医生减少勾画中的不确定性。

总之,Melvin教授总结人文大量数据的汇集能够为靶区勾画和计划设计提供以数据为基础的方法;在这个不断创新的时代,靶区勾画准确性将会提高。

Target Delineation in Head and Neck Radiotherapy: A time-honoured tradition in the era of Big Data and Big Machine

Professor Melvin demonstrated how to achieve precision target contouring in the era of big data and big machine in combination with the time-honoured tradition, so as to ensure patient survival. He pointed out that robust quality assurance is crucial in head and neck radiotherapy planning processes, and accurate target contouring is one of the key steps. Using NPC as a case example, conventional concepts on delineation of primary gross tumor volume (GTV), as well as clinical target volume show several limitations.

As recognition of anatomy and “abnormal” vs “normal” signals can be subjective, NPC primary GTV contouring is highly heterogeneous even between experienced radiation oncologists; and it is extremely labor intensive and highly depends on oncologists’ experience. So, is it possible to automate the process using deep convolutional neural networks (CNN)? A study from Sun Yan-sen University Cancer centre demonstrated that applying three-dimensional CNN (3D CNN) to automate GTV contouring on multi-parametric MR images could performed comparably to experienced radiation oncologists, outperforming 5 of the 8 radiation oncologitst. Additionally, AI-assistance helped to reduce inter-observer variation (by 2-fold) and time-taken (by 40%) substantially.

For neck CTV, current researches are focused on omitting the lower neck or treating neck to lower dose. Harnessing the power of big data, Lin et, al. marked 10651 nodes from 959 patients on a template CT scan to establish a neck nodes distribution probability map for NPC. Based on the distribution of LNs and international guidelines, they suggested 7 moderate modifications of the guidelines defined neck node levels boundaries to make it more specific to NPC.

Although a newly published international guideline on CTV delineation for NPC offers a guide to reduce ambiguity of contours between physicians, Professor Melvin pointed out that are we able to refine the CTV coverage by frequency mapping and computing tumor invasion probability of uninvaded voxels leveraging on big Data? Hence, generate probabilistic CTV by exploring machine learning methods, using large sample of GTV data, and computing tumor invasion probability of uninvaded voxels will help to further reduce uncertainties in CTV contouring

In conclusion, aggregation of large datasets will provide a data-based approach to target contouring and new age of innovation that will improve accuracy and importantly patient outcomes.

版权声明

版权属肿瘤资讯所有。欢迎个人转发分享,其他任何媒体、网站如需转载或引用本网版权所有内容,须获得授权,且在醒目位置处注明“转自:良医汇-肿瘤医生APP”。