脑网络组研究知识库

世界人口老龄化趋势不断加深，而大脑在结构和功能上的衰老进程却仍尚未得到充分地研究。磁共振成像作为非侵入式的技术，能够无创地采集脑影像数据，为解开大脑衰老的生命谜题提供了更大的可能性，倘若找到能够有效表征大脑生命周期的影像学生物标记物，便能一方面对大脑的生命周期增进认识和理解，另一方面也能够推动临床应用，使个性化治疗成为可能。为此，本文开展了基于磁共振影像的脑龄预测与大脑发展老化模式的研究。本文收集、整理和使用了2007例健康被试的结构磁共振脑影像(structural Magnetic Resonance Imaging, sMRI)数据和543例健康被试的功能磁共振脑影像(functional Magnetic Resonance Imaging, fMRI)数据作为研究的数据基础，均为大跨度数据集，所使用的数据年龄范围为18-89岁。本文所开展的研究内容可分为如下两个有所侧重又密切联系的部分：

(1)基于深度学习与脑影像的个体水平脑龄预测模型。此项工作以个体水平预测为中心，构建了基于sMRI和深度学习的脑龄预测模型并进行了验证和对比。对健康被试的数据进行了预处理，包括去头骨、偏差矫正、降噪、尺寸裁剪等，模型上结合了监督学习范式和度量学习范式，实现了多任务学习，在以三维卷积为基础的模型网络框架中融入了Inception启发的多分支结构、ResNet启发的短路结构和SENet启发的通道自适应注意力机制等技术，在增加模型网络深度和表达能力的同时减少了参数量，采取了标签平滑、权重衰减等技术作为正则化手段，提升了模型的预测精度和泛化能力，在训练时采用数据增广和AdamW自适应优化策略，并在硬件平台上实现了双卡并行。本文提出的模型具备较高的预测精度和更少的参数量，使用模型集成达到的精度为：MAE=3.24, r=0.98, p<0.001，单个模型的参数量为222万，相较使用相近样本量和年龄范围的研究，本文提出的模型表现优异，这也为下一部分的研究以及个性化治疗的临床应用奠定了基础。

(2)基于多模态脑影像探究大脑发展老化模式。此项工作从多模态、多水平角度分别研究大脑的发展和老化模式，可以分为两个角度：基于sMRI的个体水平上的大脑结构发展老化模式和基于fMRI的群组水平上的大脑功能发展老化模式。大脑结构发展老化模式方面，首先基于第一部分构建的个体水平脑龄预测模型和梯度加权类激活图(Gradient-weighted Class Activation Mapping, Grad-CAM)算法提取到sMRI样本不同脑区灰质对脑龄预测的贡献图，将样本划分年龄段对贡献图进行相关分析，从青年、中年、老年三个年龄段分别对灰质结构变化的规律以及纵向对比的灰质结构变化进行了研究，体现出大脑灰质结构差异的异质性。大脑功能发展老化模式方面，本文基于功能连接的相关性提取群组内的个体变异性作为生物标志物，并表明大脑功能的个体变异性可以反映行为表现，之后以个体变异性为基础，对样本划分年龄群组进行相关分析，在体素水平和大规模脑网络水平上对个体差异性进行了研究，发现随着年龄增长，大脑灰质功能上的个体差异在不断增加，但区域间的个体差异的不同在减小，具体地，感知网络和视觉网络等初级功能脑网络的个体变异性升高，腹侧注意网络和额顶网络等高级认知网络始终保持较高的个体变异性，且在同年龄区间内的个体变异性也均比初级功能脑网络高，这些现象揭示了功能连接个体差异的不同模式。这些发现能够增进对成年人大脑发展老化过程的基本原理的理解。

Despite our aging population, the structural and functional aging processes of the brain remain poorly understood. Magnetic resonance imaging (MRI), which acquires brain imaging data in a non-invasive manner, offers great potential for solving the mysteries of brain aging. The discovery of imaging biomarkers that can effectively characterize the brain life cycle will not only improve our understanding of the life cycle of the brain but also facilitate clinical applications and personalized treatment. We conducted a study on brain age prediction and brain development aging patterns based on structural MRI (sMRI) and functional MRI (fMRI) data obtained from 2007 and 543 healthy subjects, respectively, with the life-span age range of 18–89 years. The research carried out in this study was divided into two closely related areas:

(1) Here, we developed and validated an individual-level brain age prediction model based on deep learning and sMRI brain imaging, which was compared  with models proposed in other studies. The sMRI data were preprocessed, including skull-stripping, bias correction, noise reduction, and size cropping. The model combined supervised and metric learning paradigms to achieve multi-task learning. Based on three-dimensional (3D) convolution, Inception-inspired multi-branch structure, ResNet-inspired short-circuit structure, and SENet-inspired channel adaptive attention mechanism were incorporated into the model framework, thus reducing the amount of parameters while enhancing the depth and expressive ability of the model. Label smoothing, weight decay, and other techniques were used for regularization to improve model prediction accuracy and generalization ability. Data augmentation and AdamW adaptive optimization strategies were used for training. Dual-card parallelism was realized using the hardware platform. The model ensemble proposed in this study showed higher  prediction accuracy and fewer parameters than models proposed in other studies ( MAE=3.24, r=0.98, p<0.001, number of parameters in a single model = 2.22 million). Compared with studies using similar sample sizes and age ranges, the proposed model performed well, thus laying the foundation for the second part of the study as well as clinical application and personalized treatment.

(2) Here, we explored brain development and aging patterns from a multi-modal and multi-level perspective, i.e., sMRI-based brain structure development and aging patterns at the individual level and fMRI-based brain function development and aging patterns at the group level. In terms of brain structure, based on the above individual-level brain age prediction model and Gradient-Weighted Class Activation Mapping (Grad-CAM) algorithm, we first extracted the gray matter contribution map of different brain regions in sMRI samples to predict brain age. The contribution map of all samples was divided into age groups and correlation analysis was conducted. We investigated the regularity and longitudinal comparison of changes in gray matter structure in the young, middle-aged, and old cohorts, reflecting the heterogeneity of gray matter structure during development and aging. In terms of brain function, we extracted individual variability (IndVar) within a group based on the correlation of functional connectivity as a biomarker and showed that IndVar in brain function reflected behavioral performance. Based on IndVar, we performed voxel and large-scale brain network correlation analysis on samples divided into different age groups. Results showed that IndVar in gray matter function increased with age, but differences in IndVar among regions decreased with age. Specifically, IndVar increased in lower-level primary functional brain networks, such as the sensory and visual networks, while IndVar always remained high in higher-level cognitive networks, such as the ventral attention and frontoparietal networks. Moreover, within the same age interval, IndVar in higher-level cognitive networks was also higher than that in lower-level primary functional brain networks. These phenomena revealed distinct patterns of individual differences in functional connectivity, which should improve our understanding of the basic principles of development and aging in the adult brain.