Detection and evolution of diffusely abnormal white matter in multiple sclerosis: a deep learning approach

多发性硬化症(MS)是年轻人中最普遍的非创伤性脱髓鞘疾病。 磁共振成像(MRI)有助于MS的诊断和辅助患者的临床治疗。 除局灶性多发性硬化症外，多发性硬化症的脑部MRI上也可见弥漫性异常白质(DAWM)。 病人。虽然还不完全了解，DAWM被认为是疾病负担的预测因子，可能 出现在疾病的早期，可能是MS神经变性的标志，然而，纵向 由于缺乏对DAWM的研究，而且对DAWM的分割是人工的，使得对DAWM的演化研究变得困难 DAWM的。这项建议的主要目的是纵向研究这位女士的DAWM的发展。 目标将通过分析先前存在的1008名MS患者的纵向MRI数据来实现 参与了NIH支持的3期盲法多中心临床试验，称为CombiRx。 CombiRx数据包括多对比磁共振成像和各种临床措施。自动识别 数字水资源管理是这项提议的关键组成部分。根据我们的初步研究，深度学习(一门课程 机器学习算法)具有自动识别DAWM并估计其体积的潜力。我们会 使用大量的CombiRx MRI数据对深度学习模型进行训练、验证和测试，并进行研究 DAWM在这个MS队列中的进化。该提案有两个主要目标。在第一个目标中，我们将开发一个深层次的 基于全卷积神经网络的DAWM、灰度自动分割学习模型 在手动分割的引导下，显示正常白质和T2高信号的病变 神经影像专家。在第二个目标中，我们将分割DAWM和所有脑组织，包括局灶性病变， 在基线和CombiRx队列中所有可用的后续扫描(最长为6.5年)。时间上的变化 计算DAWM和T2病灶的体积、位置和MRI参数。我们将最终测试 DAWM是局灶性T2病变的先兆，与T2病变的分辨率相关，还是与单独的疾病相关 整个过程。如果DAWM被证明发生在疾病的早期，那么有可能更早地干预 改善了结果。同样，如果DAWM被证明与疾病活动有关，它可以作为一个客观指标 和疾病的定量测量。这样一种客观的衡量在 开发有针对性的治疗方法并评估多发性硬化症患者的治疗效果。

Multiple sclerosis (MS) is the most widespread non-traumatic, demyelinating disorder in young adults. Magnetic resonance imaging (MRI) aids in both diagnosing MS and assisting clinical management of patients. In addition to focal MS lesions, diffusely abnormal white matter (DAWM) is also seen on brain MRI in MS patients. While not understood completely, DAWM is thought to be a predictor of disease burden, possibly appears early on in the disease, and may be a marker of neurodegeneration in MS. However, longitudinal studies of DAWM are lacking, and segmentation of DAWM is manual, making it difficult to study the evolution of DAWM. The main objective of this proposal is to longitudinally study the development of DAWM in MS. This objective will be realized by analyzing preexisting longitudinal MRI data acquired on 1008 MS patients who participated in phase 3, blinded, multi-center clinical trial, referred to as CombiRx that was supported by NIH. The CombiRx data includes multi-contrast MRI and various clinical measures. Automatic identification of DAWM is a critical component of this proposal. Based on our preliminary studies, deep Learning (a class of machine learning algorithms) has the potential to automatically identify DAWM and estimate its volume. We will use the large CombiRx MRI data for training, validation, and testing of the deep learning models, and to study DAWM evolution in this MS cohort. The proposal has two major aims. In the first aim we will develop a deep learning model based on fully-convolutional neural networks for automatic segmentation of DAWM, gray matter, normal appearing white matter, and T2-hyperintense lesions guided by manual segmentation of two neuroimaging experts. In the second aim we will segment DAWM and all brain tissues, including focal lesions, at baseline and all available follow-up scans in the CombiRx cohort (up to 6.5 years). The temporal changes in volume, location, and MRI parameters of DAWM and focal T2 lesions will be computed. We will finally test whether DAWM is precursor to focal T2 lesions, associated with T2 lesion resolution, or a separate disease process altogether. If DAWM is shown to occur early on in the disease, it is possible to intervene sooner for improved outcome. Similarly, if DAWM is shown to be related to disease activity, it can serve as an objective and quantitative measure of the disease. Such an objective measurement would be highly valuable in developing targeted therapies and also in evaluating the treatment effect in MS patients.