Diagnosis of indeterminate brain lesions using MRI-based machine learning and polygenic risk models

项目摘要 2017年，每10名美国居民中就有一名以上的人接受MRI检查。其中大多数是 脑部核磁共振在45岁以上的人群中，超过1%的脑部MRI显示存在不确定的肿块病变。 脑MRI的错误解释可导致显著的医源性发病率和死亡率。比如说， 肿胀性中枢神经系统炎性脱髓鞘疾病（CNSIDD）通常是 被误诊为恶性肿瘤，甚至在病理检查之后。这会导致不适当的脑活检， 减积和放疗。虽然早期肿瘤切除与高血压患者的良好结局相关， 分级胶质瘤、观察、在替代部位活检或非手术选择通常更适合于 其他不确定的肿块病变，可能包括低级别原发性脑肿瘤、CNSIDD、CNS 淋巴瘤和脑转移。因此，为了防止医源性发病率，迫切需要可扩展且 可重复的方法，以区分CNSIDD与其他脑病变，并准确诊断脑肿瘤 在活检之前。我们最近发表了一个多基因风险模型，表明25个已知的胶质瘤生殖细胞系 风险变量可以估计胶质瘤的绝对和终生风险。这些模型的临床意义是驱动 与胶质瘤风险增加>4倍相关的生殖系变异。我们还表明， 相同的25个种系变异可以预测胶质瘤分子亚型。作为补充办法，我们 显示胶质瘤、CNSIDD、CNS淋巴瘤和脑转移瘤的成像特征不同。我们 已经成功地利用基于MRI的机器学习来预测高级别胶质瘤的分子亚型。 我们假设生殖细胞基因分型和基于MRI的机器学习都提供了一个机会， 诊断不确定的肿块病变以及在手术前预测胶质瘤分子亚型， 个性化治疗。该项目有以下三个目标：目标1：开发和验证基于MRI的 机器学习模型区分成人弥漫性胶质瘤与肿瘤性CNSIDD，CNS淋巴瘤和 原发灶不明的单发脑转移瘤。目的2：评价多基因扩增的敏感性和特异性 胶质瘤风险模型用于区分成人弥漫性胶质瘤与肿瘤性CNSIDD、CNS淋巴瘤和孤立性胶质瘤 脑转移目的3：整合多基因胶质瘤亚型模型和基于MRI的机器学习 模型预测成人弥漫性胶质瘤分子亚型，并使用前瞻性 队列。拟议的项目将通过确定早期MRI病变是否 其实是神经胶质瘤这些患者的早期确定性手术可能是治愈性的。

PROJECT SUMMARY In 2017 an MRI was performed at a rate of over one for every 10 US residents. The majority of these were brain MRIs. Indeterminate mass lesions are present on over 1% of brain MRIs in individuals over 45 years old. Misinterpretation of brain MRI can lead to significant iatrogenic morbidity and mortality. For example, tumefactive Central Nervous System Inflammatory Demyelinating Disease (CNSIDD) is commonly misdiagnosed as a malignancy, even following pathological review. This results in inappropriate brain biopsies, debulking and radiation. While early tumor resection is associated with favorable outcome in patients with high- grade glioma, observation, biopsy at an alternate site or nonsurgical options are often more appropriate for other indeterminate mass lesions that can encompass low-grade primary brain tumor, CNSIDD, CNS lymphoma and brain metastasis. Thus, to prevent iatrogenic morbidity, there is a critical need for scalable and reproducible methods to distinguish CNSIDD from other brain lesions, and to accurately diagnose brain tumors prior to biopsy. We recently published a polygenic risk model demonstrating that the 25 known glioma germline risk variants can estimate absolute and lifetime glioma risk. The clinical significance of these models is driven by germline variants that are associated with >4-fold increased risk of glioma. We have also shown that the same 25 germline variants can predict glioma molecular subtype. As a complementary approach, we have shown that imaging characteristics differ across glioma, CNSIDD, CNS lymphoma and brain metastases. We have successfully utilized MRI-based machine learning to predict the molecular subtype in high-grade glioma. We hypothesize that both germline genotyping and MRI-based machine learning provide an opportunity to diagnose indeterminate mass lesions as well as predict glioma molecular subtype prior to surgery and thus personalized treatment. The project has the following three aims: Aim 1: Develop and validate a MRI-based machine learning model to differentiate adult diffuse glioma from tumefactive CNSIDD, CNS lymphoma and solitary brain metastases of unknown primary. Aim 2: Evaluate sensitivity and specificity of the polygenic glioma risk model to differentiate adult diffuse glioma from tumefactive CNSIDD, CNS lymphoma and solitary brain metastases. Aim 3: Integrate the polygenic glioma subtype model and MRI-based machine learning model to predict adult diffuse glioma molecular subtype and validate the integrated model using a prospective cohort. The proposed project will further enhance the care of patients by determining if an early MRI lesion is actually a glioma. Early definitive surgery in these patients could be curative.