Spatio-temporal Generative Manifolds for Prediction of Immunotherapy Response

免疫疗法已被证明为治疗选择有限的癌症患者提供长期和治愈性益处，有可能彻底改变癌症治疗，并成为治愈许多疾病的综合治疗方案的重要组成部分。但是，尽管这些进展令人鼓舞，但仍然没有可靠的成像肿瘤生物标志物可以准确识别对免疫治疗效果有反应的患者或监测他们的反应以调整治疗策略，最高报告的检测率为71%。为此，放射组学是活检的定量非侵入性替代方法，可以生成表征肿瘤空间异质性的图像驱动生物标志物，并更好地识别免疫治疗反应，但缺乏预测治疗结果的统计框架。该研究计划的全球目标是开发一个用于判别流形嵌入的免疫治疗反应预测平台，使用局部保留标准减少时间域中的复杂和多模态医学成像数据。时空流形学习框架将能够恢复免疫表型特异性的放射组学签名。该研究计划将结合高度创新的理论发展和应用项目，其目标如下：（1）开发一种流形嵌入技术，使用概率深度神经网络将反应和非反应患者的医学图像映射到免疫治疗;（2）参数化流形，以量化对结果的治疗效果，处理基于梯度的优化挑战;（3）基于黎曼几何的时空流形创建治疗反应预测框架;（4）在癌症患者生物库上应用并验证所提出的方法，以将预测结果与实际结果进行比较。长期目标是提供必要的工具，用于从给定的免疫治疗癌症患者队列中生成代表性知识，并从医学成像中发现免疫系统反应和肿瘤演变的预测性生物标志物。该研究计划将利用流形和深度学习的最新进展提出新的结果预测工具。虽然以前的预测性免疫治疗反应方法是基于使用手工特征的线性回归，并且缺乏捕获免疫效应的潜在特征的可解释表示，但我们将开发一个框架，该框架将学习潜在流形的空间和时间特性。与传统的机器学习技术对高维数据的内在拓扑结构敏感相比，我们将在联合嵌入中利用新的概念，为广泛的成像和临床数据增加灵活性和多功能性。从医学的角度来看，这项研究有可能为癌症的早期检测和治疗结果的预测提供新的见解和指征。

Immunotherapy has been shown to provide long-term and curative benefits to cancer patients who have limited treatment options, with the potential to revolutionize cancer therapy and become an important part of comprehensive treatment options for curing many diseases. But while these advances are encouraging, there is still no reliable imaging tumor biomarker that can accurately identify patients responding to immunotherapy effects or monitor their response to adjust the therapeutic strategy, with highest reported detection rates of 71%. Towards this end, radiomics are quantitative non-invasive alternatives to biopsies generating image-driven biomarkers characterizing tumor spatial heterogeneity and better identify immunotherapy response, but lack the statistical framework to predict therapy outcomes.******The global objective of this research program is to develop a prediction platform for immunotherapy response from discriminant manifold embeddings, reducing complex and multi-modal medical imaging data in a temporal domain using locality preservation criterions. The spatial-temporal manifold learning framework will be able to recover radiomic signatures which are specific to immune phenotypes. The research program will incorporate highly innovative theoretical developments and applied projects with the following objectives: (1) Develop a manifold embedding technique mapping medical images of responsive and non-responsive patients to immunotherapy using probabilistic deep neural networks; (2) Parameterize manifolds to quantify therapeutic effects on outcomes, handling challenges in gradient-based optimization; (3) Create a treatment response prediction framework based on spatial-temporal manifolds anchored on Riemannian geometry; (4) Apply and validate the proposed methods on cancer patient biobanks to compare predicted results with actual outcomes.***The long-term goal is to offer the necessary tools for generating representative knowledge from a given cohort of immunotherapy cancer patients and discovering predictive biomarkers of immune system response and tumor evolution from medical imaging.******This research program will propose novel outcome prediction tools using recent advances in manifold and deep learning. While previous predictive immunotherapy response methods were based on linear regression using hand-crafted features and lacked interpretable representations of underlying features capturing immunologic effects, we will develop a framework that will learn spatial and temporal properties of underlying manifolds. Compared to traditional machine learning techniques sensitive to the intrinsic topology of high-dimensional data, we will exploit novel concepts in joint embeddings, adding flexibility and versatility capabilities to a wide range of imaging and clinical data. From a medical outlook, this research has the potential to contribute new insights and indications for possible early detection of cancer and predictors of therapeutic outcomes.