Deep Learning for Detecting the Early Anatomical Effects of Alzheimer's Disease

项目摘要 纵向的受试者内方法有可能提高灵敏度和特异性， 通过减少受试者数量和提供潜在的替代终点来提高临床试验的效率， 评估治疗效果。这些工具也有很大的潜力， 解剖建模，以更好地了解进展的时间动态。老年痴呆症 特别是，在广泛的和可能不可逆的细胞死亡之前的早期检测，对于开发 有效的治疗干预。然而，纵向工具还没有被优化用于 临床研究。挑战包括在提供每个时间点的同时减少连续扫描的噪声 相等的相对权重以避免偏倚;充分和适当地解释萎缩;以及处理 随时间变化的MRI对比度和失真。在这一建议中，我们试图改善纵向分析， 许多方法，利用现代深度学习的力量来提高准确性，使其适用于 任何类型的MRI对比，从根本上减少执行时间，以及使其可用于直接临床 应用. 为了实现这些目标，我们将采用新开发的图像合成技术来训练网络，以检测 隐藏在一组大规模“MRI”扭曲中的微小“真实”解剖学变化， 图像采集中的纵向差异，例如梯度非线性、场强和B 0失真， 和序列参数变化。变化检测网络将与深度配准级联 该网络将学习将时间扭曲分解为不感兴趣的MRI失真和感兴趣的MRI失真。 解剖效果，则扭曲场和对齐的图像都将被提供给分割网络 以确保注册不会丢失任何信息。网络将学习忽略MRI效应， 他们的刻板行为（例如B 0失真的一维性，梯度的空间平滑性 非线性）并检测细微的解剖学变化，例如心室大小的增加或心室大小的减小， 海马体积其结果将是一套强大的对比度和失真不可知的工具，突出 对临床医生的潜在疾病影响。

Project Summary Longitudinal, within-subject approaches, have the potential to increase sensitivity and specificity, improving the efficiency of clinical trials by requiring fewer subjects and providing potential surrogate endpoints to assess therapeutic efficacy. There is also great potential that these tools will enable more sophisticated anatomical modeling to better understand the temporal dynamics of progression. In Alzheimer’s Disease in particular, early detection, prior to widespread and likely irreversible cell death, is crucial for the development of effective therapeutic interventions. However, longitudinal tools have not yet been optimized for use in clinical studies. Challenges include the reduction of noise across serial scans while providing each time point equal relative weighting to avoid bias; adequately and appropriately accounting for atrophy; and handling varying MRI contrast and distortion across time. In this proposal, we seek to improve longitudinal analysis in a number of ways, leveraging the power of modern deep learning to increase accuracy, make it applicable to any type of MRI contrast, radically reduce execution time, as well as make it usable in direct clinical applications. To achieve these aims we will employ newly developed image synthesis techniques to train networks to detect small, “true” anatomical change hidden within a set of large-scale “MRI” distortions, that will capture longitudinal differences in image acquisition such as gradient nonlinearities, field strength and B0 distortions, and sequence parameter variations. The change-detection network will be cascaded with a deep registration network that will learn to decompose the temporal warp into uninteresting MRI distortions and interesting anatomical effects, then both warp fields and the aligned images will be provided to a segmentation network to ensure no information is lost by the registration. The networks will learn to ignore MRI effects based on their stereotypical behavior (e.g. the one-dimensionality of B0 distortions, the spatial smoothness of gradient nonlinearities) and to detect the subtle anatomical changes such as increasing ventricular size or decreasing hippocampal volume. The result will be a set of robust contrast-and-distortion-agnostic tools that highlight potential disease effects for clinicians.