Precision imaging for risk stratification and personalized therapy of oropharyngeal cancer

摘要 在过去的几十年里，口咽部疾病的发病率迅速增加， 癌症（OPC），其可归因于口腔人乳头瘤病毒（HPV）感染的流行。 HPV阳性OPC患者对同步放化疗反应良好， 预后优于HPV阴性患者。然而，标准治疗与 严重的毒性，可能代表许多HPV阳性疾病患者的过度治疗。 几项随机临床试验已经测试了新的去强化策略，目的是 在保持高治愈率的同时，降低毒性，提高患者的生活质量。这些试验 根据癌症分期和吸烟史招募患者。然而，目前的临床预后 这些因素相当粗糙，不能准确地预测个体水平上的疾病进展。 OPC的个性化风险适应性治疗迫切需要可靠的预后模型。到 为了解决这一未满足的需求，我们提出了定量CT特征来表征肿瘤内空间 异质性和疾病侵袭/传播，这是治疗耐药性的已知驱动因素， 疾病进展。除了基于知识的图像特征，我们还将开发互补的 数据驱动的深度学习模型，通过使用回顾性的多个 1771名患者的机构数据集。此外，我们将把成像与临床数据结合起来， 预测，并建立其有效性，通过严格的前瞻性验证，在1780例患者参加了3 随机临床试验。最后，我们将采用放射基因组学的方法来阐明生物 图像签名的基础。如果成功，所提出的模型将允许更准确的 预测预后和改善OPC的危险分层。这具有显著的治疗效果， 通过优化选择患者进行治疗去强化的影响，这将增加 未来临床试验成功的可能性，并为OPC中的精准医学铺平道路。 因为信息是从标准CT扫描中得到的，所以这将容易地集成到 目前的临床工作流程，广泛适用于低资源环境中服务不足的人群，以及 因此，这将有助于减少美国的健康差距。

ABSTRACT During the past few decades, there has been a rapid increase in the incidence of oropharyngeal cancer (OPC), which is attributable to the epidemic of oral human papillomavirus (HPV) infection. Patients with HPV-positive OPC respond well to concurrent chemoradiotherapy and have a more favorable prognosis than HPV-negative patients. However, standard treatment is associated with significant toxicity and likely represents over-treatment for many patients with HPV-positive disease. Several randomized clinical trials have tested novel deintensification strategies with the goal to reduce toxicity and improve patients’ quality of life while preserving the high cure rate. These trials enroll patients based on cancer stage and smoking history. However, current clinical prognostic factors are rather crude and do not accurately predict disease progression on an individual level. Reliable prognostic models are critically needed for personalized risk-adaptive therapy of OPC. To address this unmet need, we propose quantitative CT features to characterize intratumoral spatial heterogeneity and disease invasion/spread, which are known drivers of treatment resistance and disease progression. In addition to knowledge-based image features, we will develop complementary data-driven deep learning models to predict disease progression by using a retrospective multi- institutional dataset of 1771 patients. Further, we will integrate imaging with clinical data to improve prediction and establish their validity by rigorous prospective validation in 1780 patients enrolled in 3 randomized clinical trials. Finally, we will employ a radiogenomic approach to elucidate biological basis of the imaging signatures. If successful, the proposed models will allow more accurate prediction of prognosis and improve risk stratification of OPC. This has significant therapeutic implications by optimizing the selection of patients for treatment deintensification, which will increase the likelihood of success of future clinical trials and pave the way for precision medicine in OPC. Because the information is derived from standard CT scans, this would be readily integrated into current clinical workflow, widely applicable to underserved populations in low-resource settings, and therefore would help reduce health disparity in the US.