Model Based Deep Learning Framework for Ultra-High Resolution Multi-Contrast MRI

迫切需要灵敏的成像生物标志物来筛查高危受试者，确定 早期疾病进展，并评估对神经退行性疾病治疗的反应。 几个脑区域的萎缩是AD中确定的生物标志物，这强烈地影响了AD的预后。 与AD神经病理学相关。子野体积和皮质厚度的准确性 从目前的MRI方法估计是有限的，因为易受运动，低 空间分辨率、大脑子结构之间的低对比度和电流依赖性 图像质量的分割框架。用于标测的短运动补偿MRI方案 人脑在高空间分辨率与多种对比度，沿着与准确， 计算上有效的分割算法，迫切需要早期检测， 管理患有神经退行性疾病的受试者。 我们建议引入15分钟运动稳健的3D采集和重建 恢复全脑MRI数据的方案，具有0.2 mm各向同性分辨率， 7 T上的不同反转时间，沿着分割算法， 加速度该框架与目前依赖于MRI的方法的关键区别在于 数据1毫米分辨率，是相当显着的空间分辨率增加到0.2毫米，以及 多个广播的可用性。这一改进得益于所有领域的创新， 数据处理管道，包括采集、重建和分析。这些 基于模型的深度学习框架促进和整合了创新 （MoDL）;该框架有助于联合利用可用的先验信息， 包括运动和磁化演变模型，与卷积神经网络 从样本数据中学习解剖信息的块。成功完成本 框架将产生敏感的生物标志物，这将大大低于PET 并且不涉及辐射暴露。随着7 T临床扫描仪变得越来越普遍， 框架可以作为高风险受试者的筛查工具（例如APOE、PSEN突变） 并在随访时间较短的患者中评估进展。

Sensitive imaging biomarkers are urgently needed for screening of high‐risk subjects, determine early disease progression, and assess response to therapies in neurodegenerative disorders. The atrophy of several brain regions is an established biomarker in AD, which strongly correlates with AD neuropathology. The accuracy of subfield volumes and cortical thickness estimated from current MRI methods is limited because of the vulnerability to motion, low spatial resolution, low contrast between brain sub‐structures, and dependence of current segmentation frameworks on image quality. Short motion‐compensated MRI protocols to map the human brain at high spatial resolution with multiple contrasts, along with accurate and computationally efficient segmentation algorithms, are urgently needed tor early detection and management of subjects with neurodegenerative disorders. We propose to introduce a 15‐minute motion‐robust 3‐D acquisition and reconstruction scheme to recover whole‐brain MRI data with 0.2 mm isotropic resolution with several different inversion times on 7T, along with segmentation algorithms that are robust to acceleration. The key difference of this framework from current approaches, which rely on MRI data 1 mm resolution, is the quite significant increase in spatial resolution to 0.2 mm as well as the availability of multiple conteasts. This improvement is enabled by innovations in all areas of the data‐processing pipeline, including acquisition, reconstruction, and analysis. These innovations are facilitated and integrated by the model based deep learning framework (MoDL); this framework facilitates the joint exploitation the available prior information, including motion and models for magnetization evolution, with convolutional neural network blocks that learn anatomical information from exemplar data. The successful completion of this framework will yield sensitive biomarkers, which will be considerably less expensive than PET and does not involve radiation exposure. As 7T clinical scanners become more common, this framework can emerge as a screening tool for high‐risk subjects (e.g. APOE, PSEN mutations) and assess progression in patients with short follow‐up duration.