Optimizing MRI for Neurologic Screening using Radiologist Crowdsourcing

摘要 在过去的几十年里，诊断和治疗急性神经系统疾病的患者， 由于新的治疗方法和成像技术的兴起， 对病人进行分类治疗目前，成像信息主要由CT提供， 描绘大血管闭塞和组织灌注。MRI还提供了对比， 可从CT和远优于上级描述的组织损伤和中风模拟。在这一点上，最近的研究 使用MRI作为急诊室的一线诊断工具， 使用CT。不幸的是，MRI在急诊和筛查应用中的使用受到其自身的高度限制。 延长成像时间。这一时间增加了成本，延迟了及时治疗，并使扫描对批量和 生理运动虽然先前已经提出了各种技术来加速MRI采集， 目前正在开发一种颠覆性和范式转移的深度学习图像重建技术 提供了前所未有的扫描时间缩短。因此，这项技术具有将MRI转变为 能够快速筛查神经系统疾病的模式。然而，深度学习图像重建 需要性能指标来设置重建过程中可能丢失的信息（例如噪声、失真） 以及必须保留哪些信息（例如，对比度、分辨率、成像特征）。共同业绩 度量是重建图像的均方误差（MSE）或结构相似性（SSIM）， 然而，众所周知，这些基于工程的指标通常不能很好地反映放射学图像 质量.该项目的总体目标是开发一种放射学上最佳的、五分钟的、多对比度的深部 学习加速MRI筛查方案我们的具体目标是开发探测方法， 将放射学偏好纳入基于深度学习的MRI重建中。为了实现这一目标，我们将 开发方法学，从人类观察者对差异损坏MRI的排序中探索图像偏好 同一主题的图像。通过众包排名研究，我们的目标是调查 专家和非专家观察员之间的感知图像质量，在多个任务中，并参考 基于工程的度量。随后，来自专家放射科医师排名的数据将用于训练图像 感知模型，近似放射科医生的偏好。该模型将用于优化 多重对比神经系统筛查方案的采样模式和重建，将在 一项试点人类受试者研究，将深度学习协议与使用传统 方法.该项目的成功完成将提供一种能够及时提供 以及神经系统筛查的相关信息。它将进一步提高我们对放射成像的理解 质量感知及其在深度学习加速MRI发展中的作用。

ABSTRACT Over the past several decades, the diagnosis and treatment of patients presenting with acute neurologic symptoms has advanced tremendously due to new therapies and the rise of imaging techniques as means to triage patients for treatment. Currently, imaging information is predominately provided by CT which provides depiction of large vessel occlusions and tissue perfusion. MRI additionally provides contrasts that are simply not available from CT and a far superior depiction of tissue damage and stroke mimics. To this point, recent studies using MRI as a frontline diagnostic tool in the emergency setting have demonstrated improved outcomes over the use of CT. Unfortunately, the use of MRI in emergency and screening applications is highly limited by its extended imaging time. This time increases costs, delays timely treatment, and sensitizes the scan to bulk and physiologic motion. While a variety of techniques have been previously proposed to accelerate MRI acquisitions, disruptive and paradigm shifting deep learning image reconstruction technology is currently being developed offering unprecedented reductions in scan times. This technology thus holds potential to transform MRI into a modality capable of rapid screening for neurologic disorders. However, deep learning image reconstructions require a performance metric to set what information can be lost across the reconstruction (e.g. noise, distortions) and what information must be retained (e.g. contrast, resolution, imaging features). Common performance metrics are the mean squared error (MSE) or structural similarity (SSIM) of reconstructed images with a ground truth; though, it is well known that these engineering-based metrics are often poor reflections of radiologic image quality. The overall goal of this project is to develop a radiologically optimal, five-minute, multi-contrast deep learning accelerated MRI screening protocol. We specifically aim to develop methods for probing and incorporating radiologic preference into deep learning based MRI reconstructions. To achieve this, we will develop methodology to probe image preference from human observer ranking of differentially corrupted MRI images of the same subject. Through crowdsourced ranking studies, we aim to investigate differences in perceived image quality between expert and non-expert observers, among multiple tasks, and in reference to engineering based metrics. Subsequently, data from the expert radiologist ranking will be used to train an image perception model that approximates the radiologist’s preferences. This model will be used to optimize the sampling patterns and reconstruction for a multi-contrast neurologic screen protocol, which will be evaluated in a pilot human subject study comparing the deep learning protocol to an abbreviated protocol using traditional methods. The successful completion of this project will provide a rapid MRI method capable of providing timely and relevant information for neurologic screening. It will further improve our understanding of radiologic imaging quality perception and its role in the development of deep learning accelerated MRI.