Advanced preoperative assessment of meningiomas

项目摘要 由蛛网膜细胞产生的脑膜瘤占所有颅内肿瘤的1/3以上。虽然通常是良性的， 这些肿瘤由于肿块效应和瘤周水肿而引起临床症状。在需要 在介入治疗中，大体全切除术在可能的情况下提供了最佳结局。治疗策略是 最终通过确定手术难度和患者整体状况之间的适当平衡来决定 健康两个力学特性是手术难度的重要预测因素：肿瘤硬度和粘附性 到周围的组织。在手术前了解这些特性将使临床医生能够更准确地 评估手术风险并确定最有效的治疗策略。机械性能难以预测 通过常规成像方法，但可以通过磁共振弹性成像直接评估 (MRE)以及相关的滑动界面成像（SII）。在已发表的研究中，我们已经表明，基于MRE的刚度 估计值与脑膜瘤和垂体腺瘤中的肿瘤硬度显著相关。 此外，SII准确预测脑膜瘤和前庭神经鞘瘤的肿瘤粘附。不过， 但是要使这些发现具有临床影响力仍然存在挑战。 对于估计刚度，主要限制在于分辨率。因此，在目标1中，我们将开发一个体素- 明智的肿瘤硬度分类器。这一目标将建立在我们最近发表的基于神经网络的反演 (NNI)，它已被证明具有上级性能优于传统的直接反演在模拟和 个脑袋在目标1a中，我们将通过实现更复杂的神经网络架构来推进NNI， 为训练创造更逼真的模拟。在目标1b中，将在体模中验证这些进展， 刚度不均匀最后，在目标1c中，在我们神经外科合作者的帮助下，我们将收集 手术刚度评估的大样本。我们将使用这些评估来训练体素刚度 分类器，然后将在单独的测试集中进行验证。这一目标将导致一个地图，传达这两个 在对外科医生有临床意义的量表上预测的刚度和置信度。 SII最紧迫的局限性包括对图像的主观解释和缺乏空间 解决预测。目标2将通过开发一个体素滑动界面分类器来解决这些挑战。在 目标2a，我们将研究基于神经网络的滑动界面预测器，以添加到我们目前的方法。在 目标2b，我们将评估这种新方法是否可以改善体模实验中的预测。在目标2c中，我们将 再次利用脑膜瘤粘附的手术评估来训练和测试逐体素分类器。的 该目的的结果将是以容易解释的概率表示的肿瘤粘附图。 总之，这些目标将为神经外科医生提供临床重要信息，以改善患者的 管理更广泛地说，该项目取得的技术进步将影响整个地雷危险教育领域。

PROJECT SUMMARY Meningiomas, which arise from arachnoid cells, make up >1/3 of all intracranial tumors. While typically benign, these tumors induce clinical symptoms due to mass effect and peritumoral edema. In cases requiring intervention, gross total resection provides the best outcomes when possible. However, treatment strategy is ultimately decided by determining the proper balance between surgical difficulty and the patient's overall health. Two mechanical properties are important predictors of surgical difficulty: tumor stiffness and adherence to surrounding tissues. Knowledge of these properties before surgery would allow clinicians to more accurately assess surgical risk and identify the most effective treatment strategy. Mechanical properties are difficult to predict by conventional imaging approaches, but can be directly assessed by Magnetic Resonance Elastography (MRE) and related Slip Interface Imaging (SII). In published studies, we have shown that MRE-based stiffness estimates are significantly correlated with tumor stiffness in meningiomas and pituitary adenomas. Furthermore, SII accurately predicted tumor adherence in meningiomas and vestibular schwannomas. Still, challenges remain to make these findings clinically-impactful. For estimating stiffness, the primary limitation lies in resolution. Therefore, in Aim 1 we will develop a voxel- wise classifier of tumor stiffness. This aim will build on our recently published neural network-based inversion (NNI), which has demonstrated superior performance to conventional direct inversions in simulation and in the brain. In Aim 1a, we will advance NNI by implementing more complex neural network architectures and creating more realistic simulations for training. In Aim 1b, the advances will be validated in a phantom with inhomogeneous stiffness. Finally, in Aim 1c with the aid of our Neurosurgery collaborators, we will collect a large sample of surgical stiffness assessments. We will use these assessments to train a voxel-wise stiffness classifier, which will then be validated in a separate test set. This aim will result in a map that conveys both stiffness and confidence in the prediction on a scale that is clinically meaningful to surgeons. The most-pressing limitations in SII include the subjective interpretation of the images and the lack of spatially resolved predictions. Aim 2 will address these challenges by developing a voxel-wise slip interface classifier. In Aim 2a, we will investigate a neural network-based predictor of slip interfaces to add to our current methods. In Aim 2b, we will evaluate if this new method can improve predictions in phantom experiments. In Aim 2c, we will again leverage surgical assessments of meningioma adherence to train and test a voxel-wise classifier. The result of this aim will be a map of tumor adherence represented as an easily interpreted probability. Taken together, these aims will provide neurosurgeons with clinically-important information to improve patient management. More broadly, technical advances made in this project will impact the entire MRE field.