Virtual Biopsy with Tissue-level Accuracy in Glioma

项目摘要 这是一项生物工程研究拨款(BRG)提案，以回应PAR-19-158，以进一步开发和 使用标准验证最关键的胶质瘤分子标志物(IDH、1p/19q、MGMT)的非侵入性小组 临床MRI T2加权图像和深度学习，并将性能扩展到组织级的准确性。 目前，获得分子标记状态的唯一可靠方法是通过直接组织采样 肿瘤，需要开颅和立体定向活组织检查或大范围开放手术切除。非侵入性 以组织水平的精确度确定分子标记将是管理中的变革 胶质瘤，减少或消除与神经外科手术相关的风险和成本，加速 最终确定治疗的时间，改善患者体验，最终改善患者结果和生存时间。 深度学习等人工智能已经成为图像数据分类的一种强大方法 这可能会超过人类的表现。使用我们的新体素分类分割的初步工作 使用NIH/NCI TCIA胶质瘤数据库的方法优于任何先前的非侵入性方法 IDH、1p/19q和MGMT甲基化的测定，准确度分别为97%、93%和95%。 分别进行了分析。然而，该方法需要在TCIA之外进行验证，并且需要精确度 扩展以达到组织级别的性能。这将通过使用我们的顶级性能 体素分类框架，利用特定于标记的目标样本大小，并获得最终提升 来自深度学习艺术品校正网络。 在目标1中，我们将建立一个包含2000多个胶质瘤的数据库，其中包括来自本研究所的500名受试者，1200名受试者。 来自我们的外部合作者，以及来自TCIA的300多名受试者。我们将训练我们的体素深度学习 基于临床T2加权磁共振图像确定分子状态的分类器，目标精度为 97%。在目标2中，我们将严格评估网络的运动和噪声敏感度，并创建一个人工产物 具有以下目标的校正网络：1)在设置大量运动/噪声时恢复精度 2)即使在没有可见伪影的情况下，也进一步提高了组织级性能的准确性。在《目标3》中，我们将 部署完整的端到端临床工作流，并评估AI工具在300上的实际性能 来自我们外部合作者的预期获得性脑瘤病例和300名受试者。人工智能工具将是 可用于其他医疗中心的部署。所开发的框架还可以扩展到 以直截了当的方式添加标记。总之，这一金砖四国倡议将进一步发展、细化和 验证基于MRI的非侵入性确定最关键的胶质瘤分子标志物的方法 组织水平的准确性，以显著减少，并在许多情况下消除了立体定向活检的需要。

Project Summary This is a Bioengineering Research Grant (BRG) proposal in response to PAR-19-158 to further develop and validate a non-invasive panel of the most critical glioma molecular markers (IDH, 1p/19q, MGMT) using standard clinical MRI T2-weighted images and deep learning, and extend the performance to tissue-level accuracies. Currently, the only reliable way of obtaining molecular marker status is through direct tissue sampling of the tumor, requiring either a craniotomy and stereotactic biopsy or a large open surgical resection. Noninvasive determination of molecular markers with tissue-level accuracy would be transformational in the management of gliomas, reducing or eliminating the risks and costs associated with a neurosurgical procedure, accelerating the time to definitive treatment, improving patient experience and ultimately patient outcomes and survival time. Artificial intelligence such as deep learning has emerged as a powerful method for classification of imaging data that can exceed human performance. Preliminary work using our novel voxel-wise classification-segmentation approach with the NIH/NCI TCIA glioma database has outperformed any prior noninvasive methods for determination of IDH, 1p/19q, and MGMT methylation, achieving accuracies of 97%, 93%, and 95%, respectively. The approach however, needs to be validated beyond the TCIA and accuracies need to be extended in order to achieve tissue level performance. This will be accomplished by using our top-performing voxel-wise classification framework, leveraging marker-specific targeted sample sizes, and gaining a final boost from deep-learning artifact correction networks. In Aim 1 we will curate a database of over 2000 gliomas including 500 subjects from our institution, 1200 subjects from our external collaborators, and over 300 subjects from the TCIA. We will train our voxel-wise deep learning classifiers to determine molecular status based on clinical T2-weighted MR images with target accuracies of 97%. In Aim 2 we will rigorously evaluate the motion and noise sensitivity of the networks and create an artifact correction network with the goals of 1) recovering accuracies in the setting of large amounts of motion/noise and 2) further boosting accuracy to tissue-level performance even in the absence of visible artifact. In Aim 3 we will deploy a complete end-to-end clinical workflow and evaluate real-world live performance of the AI tool on 300 prospectively acquired brain tumor cases and 300 subjects from our external collaborators. The AI tool will be made available for deployment at other medical centers. The developed framework can also be extended to additional markers in a straightforward fashion. In summary, this BRG proposal will further develop, refine and validate a non-invasive MRI-based method for determining the most critical glioma molecular markers rivaling tissue-level accuracies to significantly reduce and in many cases eliminate the need for stereotactic biopsy.