Technology Research Laboratory, Shimadzu Corporation, Kyoto, 619-0237, Japan.
Research Center for Charged Particle Therapy, National Institute of Radiological Sciences, Chiba, 263-8555, Japan.
Br J Radiol. 2020 May 1;93(1109):20190420. doi: 10.1259/bjr.20190420. Epub 2020 Feb 28.
For real-time markerless tumour tracking in stereotactic lung radiotherapy, we propose a different approach which uses patient-specific deep learning (DL) using a personalised data generation strategy, avoiding the need for collection of a large patient data set. We validated our strategy with digital phantom simulation and epoxy phantom studies.
We developed lung tumour tracking for radiotherapy using a convolutional neural network trained for each phantom's lesion by using multiple digitally reconstructed radiographs (DRRs) generated from each phantom's treatment planning four-dimensional CT. We trained tumour-bone differentiation using large numbers of training DRRs generated with various projection geometries to simulate tumour motion. We solved the problem of using DRRs for training and X-ray images for tracking using the training DRRs with random contrast transformation and random noise addition.
We defined adequate tracking accuracy as the percentage frames satisfying <1 mm tracking error of the isocentre. In the simulation study, we achieved 100% tracking accuracy in 3 cm spherical and 1.5×2.25×3 cm ovoid masses. In the phantom study, we achieved 100 and 94.7% tracking accuracy in 3 cm and 2 cm spherical masses, respectively. This required 32.5 ms/frame (30.8 fps) real-time processing.
We proved the potential feasibility of a real-time markerless tumour tracking framework for stereotactic lung radiotherapy based on patient-specific DL with personalised data generation with digital phantom and epoxy phantom studies.
Using DL with personalised data generation is an efficient strategy for real-time lung tumour tracking.
为了实现立体定向肺部放射治疗中的实时无标记肿瘤跟踪，我们提出了一种不同的方法，该方法使用基于患者的深度学习（DL），采用个性化的数据生成策略，避免了收集大量患者数据集的需要。我们通过数字体模模拟和环氧树脂体模研究验证了我们的策略。
我们使用针对每个体模病变的卷积神经网络来开发用于放射治疗的肺部肿瘤跟踪功能，该网络通过使用来自每个体模治疗计划四维 CT 的多个数字重建射线照片（DRR）进行训练。我们通过使用各种投影几何形状生成大量训练 DRR 来训练肿瘤-骨骼区分，以模拟肿瘤运动。我们通过使用具有随机对比度变换和随机噪声添加的训练 DRR 解决了使用 DRR 进行训练和 X 射线图像进行跟踪的问题。
我们将足够的跟踪精度定义为满足等中心<1mm 跟踪误差的帧数百分比。在模拟研究中，我们在 3cm 球形和 1.5×2.25×3cm 椭圆形肿块中实现了 100%的跟踪精度。在体模研究中，我们分别在 3cm 和 2cm 球形肿块中实现了 100%和 94.7%的跟踪精度。这需要 32.5ms/帧（30.8fps）的实时处理。
我们通过数字体模和环氧树脂体模研究证明了基于患者的、具有个性化数据生成的、基于深度学习的实时无标记肿瘤跟踪框架的潜在可行性。
使用基于患者的个性化数据生成的深度学习是实时肺部肿瘤跟踪的有效策略。
