Deep learning and radiomic biomarkers for non-invasive immuno-oncology prognostication in pan-cancer patients.

癌症是世界范围内主要的死亡原因之一，其影响正在不断扩大，与世界老龄化和人口增长平行。虽然免疫肿瘤学(IO)正在给癌症治疗带来革命性的变化，但只有一小部分患者对免疫治疗有反应，而且现有的生物标志物是不够的。最近在医学成像和高通量计算方面的创新允许将数字医学图像转换为超出视觉感知的可挖掘的高维数据，即放射组学的概念。在这个项目中，我们致力于利用人工智能(AI)计算机视觉技术进行基于非侵入性成像的IO生物标记物发现，以改善患者预后和人群健康。我们的目标是发现、验证和测试新的成像生物标记物，以改善患者分层和治疗决策，以便免疫治疗最有可能受益的患者接受免疫治疗，并减少最可能不受益于免疫治疗的患者的不良事件。此外，我们的目标是补充现有的生物标记物，以提高它们的性能。最重要的是，研究创新允许临床翻译。因此，我们的项目强烈关注有效性、实用性和临床可用性。我们将把我们的模型与突变图谱等生物数据结合起来，以揭开成像生物标记物背后的潜在机制，并将使用激活映射等人工智能技术来提高我们模型的可解释性。通过这样做，我们的目标不仅是尽可能地进行最好的验证，而且让研究界和临床从业者有可能获得对医学成像领域的创新的信任，这是成功的临床翻译和可持续研究的重要基石。我们的最后一步旨在通过向科学社会提供数据和软件代码来共享数据以促进开放科学。

Cancer is one of the leading causes of death worldwide, and its impact is continuously growing, paralleling the world’s aging, and enlarging population. While Immuno-oncology (IO) is revolutionizing cancer care, only a subset of patients respond to immunotherapy, and established biomarkers are insufficient. Recent innovations in medical imaging and high-throughput computing allow the conversion of digital medical images into mineable high-dimensional data beyond visual perception, i.e., the concept of radiomics. In this project, we strive to leverage artificial intelligence (AI) computer vision technology for non-invasive imaging-based IO biomarker discovery to improve patient outcomes and population health. Our objective is to discover, validate and test novel imaging biomarkers to improve patient stratification and treatment decision-making so that immunotherapy treatment can be given to the patients who will most likely benefit from the treatment and adverse events can be reduced in patients who will most likely not benefit from immunotherapy. Further, we aim to complement existing biomarkers to improve their performance. It is of utmost importance that research innovations allow for clinical translation. Therefore, our project strongly focuses on validity, utility, and clinical usability. We will combine our models with biological data such as mutation profiles to unravel the potential mechanisms that underly the imaging biomarkers and we will use AI techniques such as activation mapping to increase the interpretability of our models. By doing so, we not only aim for the best possible validation but also to give the research community and clinical practitioners the possibility to gain trust in novel innovations in medical imaging, a crucial cornerstone for successful clinical translation and sustainable research. Our final step aims at sharing the data to promote open science by making the data and software code available to the scientific society.