Artificial Intelligence powered virtual digital twins to construct and validate AI automated tools for safer MR-guided adaptive RT of abdominal cancers

总结 磁共振成像引导的自适应放射治疗（MRgART）允许更安全的治疗， 难以治疗的腹部软组织癌症，例如发生在靠近腹部的不可手术的胰腺癌。 高度移动的和放射敏感的胃肠（GI）器官。MRgART可实现每日重新规划， 通过改善肿瘤和附近器官的可视化来检测器官形状的变化。然而，在 腹部器官在治疗分次之间和治疗分次期间显著移动， 这些组织中累积的剂量分布的信息目前还无法获得。因此，肿瘤处方 为了保守地降低辐射风险，覆盖范围仍然经常被限制在次优水平 对胃肠道器官的毒性我们假设准确估计周围移动的健康器官的剂量， 在所有部分上累积，将使得能够对整个范围进行更少保守和更有效的治疗 疾病。因此，我们将解决关键的临床需求，以确保改善局部控制并降低发病率， 局部肿瘤进展和发病率，特别是在邻近管腔GI器官的肿瘤中， 开发可靠、准确的可变形图像配准方法，以估计空间剂量 在整个处理过程中，从先前的组分累积到移动的GI管腔器官。这项建议 通过开发、严格验证和系统测量目标的增益， 利用虚拟数字双胞胎队列进行创新的深度学习累积剂量覆盖。在 目的1，我们将开发患者特异性虚拟数字双胞胎队列，模拟21种不同的时间变化 包括呼吸和消化运动的真实GI运动。双胞胎将联合收割机分析建模 广泛使用的XCAT数字幻象。在目标2中，虚拟数字双胞胎将用于优化和 严格验证我们针对GI器官的创新渐进式配准分割深度学习网络。 这种方法的关键技术新奇在于其能够执行时空变化正则化， 对大变形进行建模，这是大多数离散方法无法实现的。在目标3中，使用AI的潜在临床收益- 与具有高剂量区域保守限值的临床标准相比， 使用VDT数据集系统地模拟各种胃肠道运动。潜在影响： 已开发和确认的人工智能技术，经确认可用于真实的生理GI运动，将适用于 并将适用于其他胃肠道软组织癌症。最终，良好的可用性- 经验证的剂量累积技术可以使临床医生定量确定累积的 辐射剂量分布到腔GI器官，并适当考虑溢出辐射，从而导致 更个性化，更安全，可能更有效的放射治疗。

SUMMARY Magnetic resonance imaging-guided adaptive radiotherapy (MRgART) allows for safer treatment of otherwise difficult-to-treat soft-tissue cancers in the abdomen, such as inoperable pancreatic cancers that occur close to highly mobile and radiosensitive gastrointestinal (GI) organs. MRgART enables daily replanning to compensate for organ shape variations through improved visualization of the tumor and nearby organs. However, nearby abdominal organs move considerably between and during treatment fractions and, crucially, accurate tracking of the dose distribution accumulated in those tissues is currently unavailable. Consequently, tumor prescription coverage is still often constrained to sub-optimal levels by design to conservatively reduce the risk of radiation toxicity to GI organs. We hypothesize that accurate estimates of doses to the surrounding mobile healthy organs, accumulated over all fractions, would enable a less conservative and more effective treatment of the full extent of the disease. Hence, the key clinical need we will address, to ensure improved local control and to reduce rates of local tumor progression and morbidity, particularly in the tumors adjacent to luminal GI organs, is the development of reliably accurate deformable image registration (DIR) methods to estimate the spatial dose accumulated to the mobile GI luminal organs throughout treatment from previous fractions. This proposal addresses the key need by developing, rigorously validating, and systematically measuring the gain in target coverage with an innovative deep learning DIR dose accumulation utilizing a cohort of virtual digital twins. In Aim 1, We will develop patient-specific virtual digital twin cohorts modeling 21 different temporally varying realistic GI motions encompassing respiratory and digestive motion. The twins will combine analytical modeling with the widely used XCAT digital phantoms. In Aim 2, the virtual digital twins will be used to optimize and rigorously validate our innovative progressive registration-segmentation deep learning network for GI organs. The key technical novelty of this approach is its ability to perform spatio-temporally varying regularization to model large deformations, not possible with most DIR methods. In Aim 3, the potential clinical gain of using AI- DIR dose accumulation compared with the clinical standard with conservative limits to the high dose region will be systematically simulated with a variety of GI tract motion using the VDT datasets. Potential impact: The developed and validated AI-DIR techniques, validated for realistic physiologic GI motions, will be applicable beyond pancreatic tumors and will apply to other GI soft-tissue cancers. Ultimately, the availability of well- validated dose accumulation techniques could enable clinicians to quantitatively determine the accumulated radiation dose distribution to luminal GI organs and appropriately account for the spillover radiation, thus leading to more personalized, safer, and possibly more effective radiation treatments.