阿尔茨海默病(Alzheimer’s Disease, AD)是一种进行性和不可逆的神经退行性疾病，以记忆丧失和认知障碍为特征。如今全球约有9000万人患有AD，预计到2050年，AD患者人数将达到1.15亿。虽然可以延缓AD进展，但是目前缺乏有效的治疗方法来阻止或逆转AD。因此，尽早诊断筛查出AD病症对于患者病情的预防和干预治疗至关重要。

AD的临床诊断过程可以通过腰椎穿刺检测脑脊液(CSF)中的特定生物标志物进行诊断，然而这是一种侵入性检查，给患者带来一定风险。AD进展的另一种常见生物学标记是脑结构的形态学改变，随着神经影像技术的发展，其对AD的准确诊断和早期发现具有不可替代的价值。磁共振成像(magnetic resonance imaging，MRI)可以无创地捕捉大脑内部结构和萎缩，帮助我们了解与AD相关的大脑解剖和功能变化，特别是T1加权成像(T1-weighted imaging, T1WI)提供了关于脑组织的内部解剖结构和形态学的详细信息，可以检测和跟踪AD脑萎缩的演变。事实上，AD诊断的显著特征之一是颞叶萎缩，尤其是海马体、杏仁核等特定皮层下结构的萎缩。

精准的脑解剖结构分割为健康人群与脑疾病人群的队列研究奠定基础，尤其是为后续特征提取、分析以及疾病分类模型的构建具有重要作用。此外，对于给定的数据集群体可以构建特定的人脑MRI模板，以提供健康人群和疾病人群的标准化参考，进行精确的神经解剖学定位、结构和功能比较，为神经科学发展和临床研究提供依据。因此本文的研究思路主要分成以下两个部分：（1）利用图像配准算法进行标签迁移以实现脑结构分割。（2）构建脑模板生成网络，生成样本群体的脑模板。

在第一部分工作中，本文提出了一种基于交叉双流的多尺度注意力特征融合网络，命名为MAFF-Net，用于脑图像微分同胚配准。首先利用交叉双流网络推断图像对之间的相互映射关系，并通过引入注意力机制融合多尺度特征的高低语义信息，最后利用微分同胚配准增强形变场的连续性和全局平滑性提高配准质量。在自采集、OASIS-AD与OASIS-Health数据集的实验结果显示，MAFF-Net算法在三个测试集上解剖结构Dice相似性系数均值分别为83.2%、85.3%和86.5%，负雅可比行列式体素比例均值为0.027%、0.192%和0.089%，Recall均值为92.4%、90.9%和92.0%，ASD均值为0.447mm、 0.387mm和0.345mm，除Recall外其余指标均优于对比算法。本文进一步利用OASIS数据集的分割结果，分析大脑皮质、海马体和杏仁核的体积和表面积与年龄的变化关系，探究这些脑结构的萎缩程度与AD的密切关系。

在第二部分工作中，本文提出了一种脑模板构建网络，命名为Template-Net。该网络是在MAFF-Net的基础上，利用脑模板作为中间态桥梁，加入个体间的相似性损失函数，构建更加精准的配准网络，同时利用给定的数据集生成精确脑模板图像。在OASIS-Health与OASIS-AD数据集上的实验结果显示，本文设计的Template-Net相比于MAFF-Net在各项指标上均取得了更优的结果。本文进一步对构建的健康人和AD脑模板进行脑结构形态学对比，以便探究AD相关的影像标记物，更准确地评估疾病群体与健康人群的差异性。

Alzheimer's disease (AD) is a progressive and irreversible neurodegenerative disease characterized by memory loss and cognitive impairment. About 90 million people worldwide have AD today, and the number of people with AD is expected to reach 115 million by 2050. Although the progression of AD can be delayed, there are currently no effective treatments to stop or reverse AD, therefore, early diagnosis and screening of AD is essential for the prevention and intervention of the condition.

The clinical diagnostic procedure for AD can be performed by a lumbar puncture to detect specific biomarkers in the cerebrospinal fluid (CSF). However, this is an invasive test that poses risks to the patient. Another common biological marker of AD progression is morphological changes in brain structure, and with the development of neuroimaging technology, it is of irreplaceable value for accurate diagnosis and early detection of AD. Magnetic resonance imaging (MRI) can noninvasively capture the internal structure and atrophy of the brain, helping us to understand the anatomical and functional changes in the brain associated with AD. In particular, T1-weighted imaging (T1WI) provides detailed information about the internal anatomical structure and morphology of brain tissue, allowing the detection and tracking of the evolution of brain atrophy in AD. Indeed, one of the distinguishing features of AD diagnosis is temporal lobe atrophy, especially of specific subcortical structures such as hippocampus and amygdala.

Accurate segmentation of brain anatomy lays the foundation for cohort studies of healthy and brain disease populations, especially for subsequent feature extraction, analysis, and construction of disease classification models. In addition, for a given data set population specific human brain MRI templates can be constructed to provide a standardized reference between healthy and diseased populations for accurate neuroanatomical localization, structural and functional comparisons for neuroscience development and clinical research. Therefore, the research idea of this paper is divided into the following two parts: (1) Using image registration algorithm for label migration to achieve brain structure segmentation. (2) Construction of a brain template generation network to generate brain templates of the sample population.

In the first part of the work, an intersected dual stream based multiscale attentional feature fusion network, named MAFF-Net, is proposed in this paper for diffeomorphic registration of brain images. Firstly, the intersected dual stream network is used to infer the mutual mapping relationship between image pairs, and fuse the high and low semantic information of multiscale features by introducing the attention mechanism, and finally, the diffeomorphic registration is used to enhance the continuity and global smoothness of the deformation field to improve the alignment quality. The experimental results on the inhouse, OASIS-AD and OASIS-Health datasets show that the MAFF-Net algorithm has mean values of 83.2%, 85.3% and 86.5% for the anatomical structure Dice similarity coefficient, mean values of 0.027%, 0.192% and 0.089% for the negative Jacobi determinant voxel ratio, mean values of 92.4%, 90.9%, and 92.0% for Recall, and mean values of 0.447 mm, 0.387 mm, and 0.345 mm for ASD on the three test sets respectively. All metrics except Recall were better than the comparison algorithm. This paper further uses the segmentation results of the OASIS dataset to analyze the volume and surface area of the cerebral cortex, hippocampus, and amygdala in relation to age, and to explore the close relationship between the degree of atrophy of these brain structures and AD.

In the second part of the work, a brain template construction network named Template-Net is proposed in this paper, which is based on MAFF-Net, using brain templates as an intermediate state bridge and adding the similarity loss function between individual images to construct a more accurate alignment network, while generating accurate brain template images using the given dataset. The experimental results on OASIS-Health and OASIS-AD datasets show that the Template-Net designed in this paper achieves better results in all metrics compared to MAFF-Net. In this paper, we further compare the morphological brain structures of the constructed healthy and AD brain templates in order to explore the AD-related imaging markers and more accurately assess the variability between the disease group and the healthy population.