Generalizable quantitative imaging and machine learning signatures in glioblastoma, for precision diagnostics and personalized treatment: the ReSPOND consortium

摘要 神经肿瘤学中的磁共振成像（MRI）方法的当前状态提供了很大的帮助。 提供丰富的结构、生理和代谢特征表征的潜力- 脑肿瘤，特别是神经胶质瘤，是复杂和高度异质性的癌症。 特别是胶质母细胞瘤（GBM），预后很差，中位总生存期（OS）小于 15个月，自Stupp方案实施以来的过去15年中， 介绍许多实验性治疗正在进行中;然而，OS在很大程度上仍然存在。 停滞不前改善这一结果的一些障碍是：1）疾病异质性， 这使得在1期甚至2期试验中难以检测治疗效果， 用于个性化，而不是一刀切的治疗策略; 2）用于肿瘤的方法 基于大小、增强、灌注和扩散特性的表征相对 粗糙并且不能充分利用成像数据的丰富性或其空间异质性。定量 在过去十年中开发的动态成像和机器学习（QIML）方法已经表明， 解剖GBM的空间、时间和患者间异质性的巨大潜力; 影像学和分子特征之间的关系;为提供个性化 临床结果的预测;以及利用数据中微妙的多参数关系 为了检测肿瘤周围浸润或区分治疗相关的变化，即，假性进展 (PsP)肿瘤复发。我们的团队一直处于QIML的最前沿，重点是 a）获得依赖于多参数信号、纹理参数、形状 性质、从图谱配准导出的空间模式和肿瘤生长的生物物理模型， 以及B）使用机器学习将这样的成像特征整合到临床诊断的预测器中， 结果、肿瘤周围浸润的早期复发、PsP和GBM的放射学亚型。 尽管QIML方法很有前途，但它们有一个众所周知的局限性：它们可能过度拟合特定的 数据集，它们是从其中派生的，并可能显示在现实生活中的再现性差， 可变扫描仪类型和成像协议的条件。在本提案中，我们旨在利用 最近成立了ReSPOND（用于精确诊断的放射组学特征）联盟， 整合、协调和分析来自全球14个中心的4，578个数据集， 更适当地培训和交叉验证QIML工具，以获得更广泛的通用性。该联盟 将产生一个前所未有的数据库，其中包含各种精心协调的MRI集， 临床措施，旨在为社区提供稳健和可重复的QIML模型 有助于精确诊断和个性化治疗这种可怕的脑癌。

Abstract The current state of magnetic resonance imaging (MRI) methods in neurooncology offers great potential for providing rich characterizations of structural, physiological, and metabolic character- istics of brain tumors, especially gliomas, which are complex and highly heterogeneous cancers. Glioblastoma (GBM), in particular, has a grim prognosis, with median overall survival (OS) less than 15 months with relatively little improvement in the past 15 years since the Stupp protocol was introduced. Many experimental treatments are being pursued; however, OS has largely remained stagnant. Some of the obstacles in improving this outcome have been 1) disease heterogeneity, which both renders it difficult to detect treatment effects in Phase 1 or even Phase 2 trials, and calls for personalized, rather than one-size fits-all, treatment strategies; 2) methods used for tumor characterization based on size, enhancement, perfusion and diffusion properties are relatively crude and don't fully leverage the richness of imaging data or their spatial heterogeneity. Quanti- tative imaging and machine learning (QIML) methods developed in the past decade have shown great potential for dissecting the spatial, temporal and inter-patient heterogeneity of GBM; for discoveringrelationships between imaging and molecular characteristics ; foroffering personalized predictions of clinical outcome; and for leveraging subtle multi-parametric relationships in the data to detect peri-tumoral infiltration or distinguish treatment related changes, i.e., pseudo-progression (PsP), from true tumor recurrence. Our group has been at the forefront of QIML, with emphasis on a) obtaining rich imaging phenotypes relying on multi-parametric signals, texture parameters, shape properties, spatial patterns derived from atlas registration, and biophysical models of tumor growth, and b) integrating such imaging signatures using machine learning into predictors of clinical outcome, early recurrence from peri-tumoral infiltration, PsP, and radiologic subtypes of GBM. Despite their promise, QIML methods have a notorious limitation: they might overfit specific datasets from which they are derived, and might display poor reproducibility under real-life conditions of variable scanner types and imaging protocols. In this proposal we aim to leverage the recently formed ReSPOND (Radiomics Signatures for PrecisiON Diagnostics) consortium, to integrate, harmonize, and analyze 4,578 datasets from 14 centers around the world, and hence more appropriately train and cross-validate QIML tools for a wider generalizability. This consortium will generate an unprecedented database of diverse and carefully harmonized sets of MRI and clinical measures, and aims to provide the community with robust and reproducible QIML models contributing to precision diagnostics and personalize treatment for this dreaded brain cancer.