Quantitative imaging phenotypic classifier for distinguishing radiation effects from tumor recurrence in Glioblastoma

摘要：美国每年有超过14，000例胶质母细胞瘤（GBM）患者接受颅外联合治疗。 手术、化疗和放疗作为他们侵袭性癌症的标准治疗。不幸的是，约40%的 这些患者将在化疗-放疗后的后续MRI扫描（T1w，T2w， FLAIR）。GBM肿瘤管理中的一个重大挑战是将这些病变鉴别为肿瘤 复发或良性治疗相关放射效应（TRRE）。这些条件相互模仿，临床上 和放射学检查。不幸的是，在缺乏可靠的诊断工具的情况下，TRRE患者将经历 不必要且可避免的侵入性立体定向脑活检（St-Bx），以确认无疾病。 然而，即使是侵入性St-Bx也具有85 - 90%的准确性，这是由于与获得 活检组织可能不代表潜在的疾病病理学。因此，建筑 非侵入性决策支持工具产生的诊断准确性不劣于St-Bx，代表了 这是一种有吸引力的解决方案，可避免良性辐射效应患者不必要的颅内St-Bx。 我们的团队已经开发了一种新的基于图像的复发风险分类器（IRRisC），使用常规MRI扫描， 在n = 58项研究中，该方法在区分肿瘤复发与TRRE方面的准确性为85%。我们 IRRisC特征初始集合包括钆（Gd）-T1w MRI上的梯度取向的无序，其具有 与TRRE相比，它在肿瘤复发方面显着更高。有趣的是，我们最近 还表明，为男性和女性构建单独的分类器， GBM存活的预后与"所有参与者"模型相比。在这个R01项目中，我们寻求进一步改善 并通过扩展我们的初始特征集（使用Gd-T1w MRI）来验证IRRisC的准确性，以包括（1） 来自解剖（T2w，FLAIR）和功能MR序列（灌注）的附加特征，（2）一类新的 来自"正常"脑实质的生物物理变形属性，以及（3）性别特异性模型的构建 利用GBM中的性别二态性，用于区分肿瘤复发与TRRE。克服局限性 以前的工作涉及小样本和缺乏组织病理学验证，我们的工作将利用最大的 多机构组织病理学证实的队列，直至n = 470项TRRE和肿瘤研究的日期 复发，以协调和验证IRRisC。此外，我们将建立我们的IRRisC的生物学基础 通过评估其与TRRE和肿瘤复发的组织病理学标志的相关性来评估其特征。最后， IRRisC将在3个临床研究中心的机器阅读器研究中作为决策支持进行验证。成功标准 对于IRRisC，它将（a）不劣于St-Bx的准确度（~85 - 90%），以及（B）识别不超过50% TRRE患者患癌症的概率。这些标准将确保IRRisC在临床上可作为 强大而可靠的分类器，避免了至少50%的患者不必要的颅内活检， TRRE，同时也保持了癌症复发的高真阳性率。

ABSTRACT: Over 14,000 Glioblastoma (GBM) patients annually in the US undergo a combination of cranial surgery, chemotherapy, and radiation as standard treatment for their aggressive cancer. Unfortunately, ~40% of these patients will be identified with a suspicious lesion on a post-chemo-radiation follow up MRI scan (T1w, T2w, FLAIR). A significant challenge in the management of GBM tumors is the differentiation of these lesions as tumor recurrence or benign treatment-related radiation effects (TRRE). These conditions mimic each other, clinically and radiographically. Unfortunately, in the absence of reliable diagnostic tools, patients with TRRE will undergo an unnecessary and avoidable invasive stereotactic brain biopsy (St-Bx) for confirmation of disease absence. However, even the invasive St-Bx has an accuracy of 85-90% due to sampling errors associated with obtaining a biopsy tissue which may not be representative of the underlying disease pathology. Consequently, building non-invasive decision support tools which yield a diagnostic accuracy that is non-inferior to St-Bx, represents an attractive solution for obviating unnecessary intra-cranial St-Bx in patients with benign radiation effects. Our group has developed a new Image-based Recurrence Risk Classifier (IRRisC) using routine MRI scans, that has demonstrated an accuracy of 85% in distinguishing tumor recurrence from TRRE, on n=58 studies. Our initial set of IRRisC features comprise disorder in gradient orientations on Gadolinium (Gd)-T1w MRI which have been shown to be significantly higher in tumor recurrence compared to TRRE. Interestingly, we have recently also demonstrated that construction of separate classifiers for males and females yielded significantly improved prognosis of GBM survival compared to an ‘all-comers’ model. In this R01 project, we seek to further improve and validate the accuracy of IRRisC by expanding our initial feature set (using Gd-T1w MRI) to include (1) additional features from anatomical (T2w, FLAIR) and functional MR sequences (perfusion), (2) a new class of biophysical deformation attributes from “normal” brain parenchyma, and (3) construction of sex-specific models to exploit sexual-dimorphism in GBM, for distinguishing tumor recurrence from TRRE. Overcoming limitations of previous work pertaining to small samples and lack of histopathological validation, our work will utilize the largest multi-institutional histopathologically confirmed cohort till date of n=470 studies of TRRE and tumor recurrence, to harmonize and validate IRRisC. Further we will establish the biological underpinning of our IRRisC features by evaluating their association with histopathological hallmarks of TRRE and tumor recurrence. Finally, IRRisC will be validated as decision support in a machine-reader study at 3 clinical sites. Criteria for success for IRRisC is that it will (a) be non-inferior to the accuracy of St-Bx (~85-90%), and (b) identify no more than 50% of patients with TRRE as having cancer. These criteria will ensure that IRRisC is clinically actionable as a robust and reliable classifier, by obviating at least 50% of unnecessary intra-cranial biopsies in patients with TRRE, while also maintaining a high true positive rate for cancer recurrence.