Measures of change in hippocampal volume derived from longitudinal MRI are a well-studied biomarker of disease progression in Alzheimer's disease (AD) and are used in clinical trials to track therapeutic efficacy of disease-modifying treatments. However, longitudinal MRI change measures based on deformable registration can be confounded by MRI artifacts, resulting in over-estimation or underestimation of hippocampal atrophy. For example, the deformation-based-morphometry method ALOHA (Das et al., 2012) finds an increase in hippocampal volume in a substantial proportion of longitudinal scan pairs from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, unexpected, given that the hippocampal gray matter is lost with age and disease progression. We propose an alternative approach to quantify disease progression in the hippocampal region: to train a deep learning network (called DeepAtrophy) to infer temporal information from longitudinal scan pairs. The underlying assumption is that by learning to derive time-related information from scan pairs, the network implicitly learns to detect progressive changes that are related to aging and disease progression. Our network is trained using two categorical loss functions: one that measures the network's ability to correctly order two scans from the same subject, input in arbitrary order; and another that measures the ability to correctly infer the ratio of inter-scan intervals between two pairs of same-subject input scans. When applied to longitudinal MRI scan pairs from subjects unseen during training, DeepAtrophy achieves greater accuracy in scan temporal ordering and interscan interval inference tasks than ALOHA (88.5% vs. 75.5% and 81.1% vs. 75.0%, respectively). A scalar measure of time-related change in a subject level derived from DeepAtrophy is then examined as a biomarker of disease progression in the context of AD clinical trials. We find that this measure performs on par with ALOHA in discriminating groups of individuals at different stages of the AD continuum. Overall, our results suggest that using deep learning to infer temporal information from longitudinal MRI of the hippocampal region has good potential as a biomarker of disease progression, and hints that combining this approach with conventional deformation-based morphometry algorithms may lead to improved biomarkers in the future.

来自纵向 MRI 的海马体积变化测量是阿尔茨海默病 (AD) 疾病进展的经过充分研究的生物标志物，并用于临床试验中以跟踪疾病缓解治疗的疗效。然而，基于变形配准的纵向 MRI 变化测量可能会受到 MRI 伪影的干扰，导致海马萎缩的高估或低估。例如，基于变形的形态测量方法 ALOHA（Das 等人，2012）发现，阿尔茨海默病神经影像计划 (ADNI) 研究中的大部分纵向扫描对中海马体积有所增加，这是出乎意料的，因为海马体积灰质随着年龄和疾病进展而丧失。我们提出了另一种量化海马区域疾病进展的方法：训练深度学习网络（称为 DeepAtropy）以从纵向扫描对推断时间信息。基本假设是，通过学习从扫描对中获取与时间相关的信息，网络隐式学习检测与衰老和疾病进展相关的渐进变化。我们的网络使用两种分类损失函数进行训练：一种用于测量网络正确排序来自同一主题的两次扫描（以任意顺序输入）的能力；另一种用于衡量网络对同一主题的两次扫描进行正确排序的能力。另一个衡量正确推断两对相同受试者输入扫描之间的扫描间间隔比率的能力。当应用于训练期间未见过的受试者的纵向 MRI 扫描对时，DeepAtropy 在扫描时间排序和扫描间间隔推理任务中比 ALOHA 实现了更高的准确性（分别为 88.5% vs. 75.5% 和 81.1% vs. 75.0%）。然后，对源自深度萎缩的受试者水平随时间变化的标量测量进行检查，作为 AD 临床试验中疾病进展的生物标志物。我们发现，在区分 AD 连续体不同阶段的个体群体方面，该措施的表现与 ALOHA 相当。总体而言，我们的结果表明，使用深度学习从海马区的纵向 MRI 推断时间信息具有作为疾病进展的生物标志物的良好潜力，并暗示将这种方法与传统的基于变形的形态测量算法相结合可能会导致改进的生物标志物未来。