Real-time radiation dose reconstruction for improved radiation therapy

导读：大约70%的癌症患者使用放射治疗。然而，目前使用放射治疗的治疗有两个主要弱点：1)它还没有提出一个很好的解决方案来解释肿瘤在治疗过程中的运动，2)它假设实际传递的辐射剂量等于计划剂量。在这个提议中，我们计划使用治疗辐射束本身来解决这两个弱点。在治疗过程中测量出患者的辐射量将使我们能够1)对肿瘤目标成像，并进行实时校正，以跟踪光束到移动的肿瘤，2)重建实际的3D交付给患者的辐射剂量，而不假设它等于计划剂量。方法：最近开发的人工智能（AI）方法提供了提供极快的图像分析的承诺，同时仍然保持高度准确。我们将训练一个人工智能网络来识别患者放射治疗期间获得的x射线图像中存在的患者产生的散射x射线。这种训练需要大量的x射线图像，其中x射线散射量是确切已知的。我们将利用我们在x射线计算方面的专业知识，包括我们之前开发的非常精确的方法，但对于实时来说太慢了，来生成这些图像。一旦人工智能网络得到正确训练，它将实时提供准确的患者x射线散射信号估计。从x射线图像中去除x射线散射信号将使图像更清晰，这有利于肿瘤的跟踪。利用这些更清晰的图像，我们将训练第二个人工智能网络，以实时识别和跟踪x射线图像中的肿瘤运动。我们将从一个被称为“幻影”的假肺癌患者开始，在那里我们可以精确地控制塑料肿瘤的运动，然后使用更真实的患者数据。最后，我们将实现两个人工智能网络一起实时跟踪肿瘤运动。在对肿瘤进行实时追踪的同时，我们将使用更清晰的图像来计算实际传送到幻体的辐射剂量。一旦在幻影上成功演示，它将准备转化为患者使用。意义：利用治疗束在治疗过程中实时校正肿瘤运动，实时准确计算实际给病人的剂量，解决当前放疗的两大弱点。总之，这些进步将改善实时肿瘤跟踪应用，并通过实时错误检测提高患者安全性。这里开发的创新方法有可能改善全世界放射治疗单位的癌症治疗。

Introduction: Radiation is used to help treat about 70% of all cancer patients. However, current treatment using radiotherapy has two major weaknesses - 1) it has yet to come up with a good solution to account for the motion of the tumour during treatment, and 2) it assumes the actual delivered radiation dose is equal to the planned dose. In this proposal, we plan to use the treatment radiation beam itself to address both of these weaknesses. Measuring the amount of radiation exiting the patient during the treatment will allow us to both 1) image the tumour target and perform real-time correction to track the beam to the moving tumour and 2) reconstruct the actual 3D delivered radiation dose to the patient, without assuming it is equal to the planned dose. Methods: Recently developed artificial intelligence (AI) methods offer the promise of providing extremely fast image analysis while still remaining highly accurate. We will train an AI network to recognize patient-generated scattered x-rays which are present in x-ray images acquired during patient radiation treatment. This training requires a large number of x-ray images where the amount of x-ray scatter is known exactly. We will use our expertise in x-ray calculations, including very accurate methods we developed previously but which are too slow for real-time, to generate these images. Once the AI network is correctly trained, it will provide accurate patient x-ray scatter signal estimates in real-time. The removal of the x-ray scatter signal from the x-ray images will result in clearer images, which is beneficial for tumour tracking. Using these clearer images, we will train a second AI network to identify and track tumour movement in the x-ray images in real-time. We will start by working with a pretend lung cancer patient called a `phantom', where we can exactly control the motion of a plastic tumour, then progress to use more-realistic patient data. Finally, we will implement the two AI networks together to track tumour motion in real-time. While the tumour is being tracked in real-time, we will use the clearer images to calculate the radiation dose that was actually delivered to the phantom. Once successfully demonstrated on the phantom, it will be ready to be translated to patient use. Significance: This research program will address two major weaknesses in current radiotherapy by using the treatment beam to correct for tumour motion in real-time during treatment, and to accurately calculate the actual delivered dose to the patient in real-time. Together, these advances will improve real-time tumour tracking applications and increase patient safety via real-time error detection. The innovative methods developed here have the potential to improve the treatment of cancer on radiotherapy units worldwide.