脑网络组研究知识库

对大脑发育、老化过程的形态学和功能的研究对于以精神分裂症、阿兹海默病、帕金森症为代表的神经退行性疾病的诊断和治疗有着重要作用。大脑的发育、老化异常会导致大脑的年龄与个体实际年龄不相同，大脑年龄的估计值可以初步而直观地反映大脑衰老的程度。而核磁共振成像技术，特别是T1加权图像，能够提供分辨率较高的三维影像，适合于对大脑内部组织结构的观察研究。因此，基于结构磁共振影像的大脑年龄的研究具有非常重要的科学意义与临床应用价值。同时，由于中国汉族人群与国外人群的脑组织形态具有一定差别，因而基于国外数据集得到的脑龄估计模型无法直接应用于国内人群，所以本文提出构建汉族人群脑龄估计模型的研究。

基于传统机器学习方法的大脑年龄估计算法，需要对大脑结构磁共振数据进行组织分割、数据降维等复杂预处理工作，且依赖于手工设计的特征，导致算法复杂度高且误差较大。己有的基于深度学习的大脑年龄估计模型仅由几个简单的卷积层堆叠而成，拟合能力不强，估计的误差较大。针对这些问题，本文在深度学习框架下，针对结构磁共振影像，设计深度卷积网络，用于脑龄预测。

本文设计了一个结合T1w结构像与深度卷积网络的脑龄预测模型。该模型中融入了NIN网络、Res Net、Inception Net网络的优势，在增加网络深度的同时，尽可能的减少模型的参数量。针对于脑龄预测任务，提出了使用均方误差与交叉熵结合的损失函数，通过多任务学习的模式使得模型在预测脑龄时具备提供预测结果置信度的能力。在模型训练的过程中，使用数据同步并行的方式提高网络训练的速度，并使用提前终止策略结束网络的训练。在模型验证环节，使用10次10折交叉验证和独立数据验证等方法验证模型的脑龄预测精度与泛化能力。验证结果显示，本文提出的模型具备很高预测精度和稳定性，模型在训练集上通过交叉验证得到的预测精度为：MAE=3.03(±2.04)，c=0.887 (p = 3.34*10^-6)，在独立数据集上验证得到的预测精度为：MAE= 3.51 (± 2.55)，c=0.85 (p = 2.49*10^-5)。与现有模型相比，本文提出的模型表现优异。最后使用Grad-CAM可视化方法将对脑龄预测有用的特征反向映射到输入空间，定位对脑龄预测用于的脑区，发现12岁和16岁对应的激活热图在颞上回、额下回、额上回、顶下小叶区域具有较大差异（p<0.01），16岁与20岁对应的激活热图在额中回、颞下回、中央后回区域具有较大差异（p<0.01）。该发现表明，发育过程（12-20岁）有一定的脑区迁移轨迹，激活程度最大的空间位置随着年龄的变化会发生改变。

其次，在上述工作基础之上，本文研究了脑龄估值差在精神分裂症中的应用。统计分析发现，精神分裂症患者组的脑龄估值差为3.2岁，与健康被试的脑龄估值差之间有显著差异（t=4.0524，p=0.0001）。使用Grad-CAM方法对于模型的可解释性进行分析，发现精神分裂症患者组的扣带回，额叶上回激活程度与健康对照组具有显著差异（p<0.01）。由此说明了脑龄估值差能够作为精神分裂症的简单直观的疾病标记物，用于疾病早期诊断和疾病治疗。

综上，本文设计了一个基于汉族人群的脑龄预测模型，该模型的预测性能优异，同时研究了脑龄估值差在精神分裂症中的临床应用价值。本文的研究内容为脑科学研究大脑发育老化奠定了脑龄预测模型基础，并为研究脑龄估值差的应用提供了思路与方向。

The study of the morphology and function of the brain development and aging process has an important role in the diagnosis and treatment of neurodegenerative diseases represented by schizophrenia, Alzheimer's disease, and Parkinson's disease. Abnormal brain development and aging will cause the age of the brain to be different from the actual age of the individual. The estimated value of brain age can initially and intuitively reflect the degree of brain aging. MRI technology, especially T1-weighted images, can provide high-resolution three-dimensional images, suitable for the observation and research of the internal tissue structure of the brain. Therefore, the study of brain age based on structural magnetic resonance imaging has very important scientific significance and clinical application value. At the same time, due to the difference in brain tissue morphology between the Chinese Han population and the foreign population, the brain age estimation model based on foreign data sets cannot be directly applied to the domestic population, so this paper proposes a study to construct the Han population brain age estimation model.

Brain age estimation algorithms based on traditional machine learning methods require complex preprocessing tasks such as tissue segmentation and data dimensionality reduction of brain structural magnetic resonance data, and rely on hand-designed features, resulting in high algorithm complexity and large errors. The existing brain age estimation model based on deep learning is only formed by stacking a few simple convolutional layers. The fitting ability is not strong, and the estimation error is large. In response to these problems, under the framework of deep learning, this paper designs a deep convolutional network for structural magnetic resonance images for brain age prediction.

In this paper, a brain age prediction model combining T1w structural image and deep convolutional network is designed. The model incorporates the advantages of the NIN network, Res Net, and Inception Net network. While increasing the depth of the network, the number of parameters of the model is reduced as much as possible. For the brain age prediction task, a loss function using a combination of mean square error and cross entropy is proposed. The multi-task learning model enables the model to provide confidence in the prediction results when predicting brain age. In the process of model training, data synchronization is used to increase the speed of network training, and an early termination strategy is used to end the network training. In the model verification process, 10 times 10-fold cross-validation and independent data verification are used to verify the prediction accuracy and generalization ability of the model. The verification results show that the model proposed in this paper has high prediction accuracy and stability. The prediction accuracy obtained by cross-validation on the training set of the model is: MAE = 3.03 (± 2.04), c = 0.887 (p = 3.34 * 10^-6), The prediction accuracy verified on an independent data set is: MAE= 3.51 (± 2.55)，c=0.85 (p = 2.49*10^-5). Compared with the existing models, the model proposed in this paper performs well. Finally, Grad-CAM visualization method is used to reversely map the input space of features useful for brain age prediction, locate the brain area used for brain age prediction, and find that the activation heat maps corresponding to 12 and 16 years old are in the upper temporal gyrus and frontal The areas of the next gyrus, upper frontal gyrus, and inferior parietal lobule have large differences (p <0.01), and the activation heat maps corresponding to 16 and 20 years old have large differences in the frontal gyrus, inferior temporal gyrus, and central posterior gyrus (p <0.01). This finding indicates that there is a certain migration trajectory of brain regions during development (12-20 years old), and the spatial location with the greatest degree of activation will change with age.

Secondly, on the basis of the above work, this paper studies the application of difference in brain age estimation in schizophrenia. Statistical analysis found that the difference in brain age estimates of the schizophrenia patient group was 3.2 years old, and there was a significant difference between the difference in brain age estimates of healthy subjects (t = 4.0524, p = 0.0001). The Grad-CAM method was used to analyze the interpretability of the model, and it was found that the degree of cingulate gyrus and upper frontal gyrus activation in the schizophrenic patient group was significantly different from that in the healthy control group (p <0.01). This shows that the difference in brain age estimation can be used as a simple and intuitive disease marker for schizophrenia for early diagnosis and treatment of disease.

In summary, this paper designs a brain age prediction model based on the Han population, which has excellent prediction performance, and at the same time studies the clinical application value of poor brain age estimation in schizophrenia. The research content of this article lays the foundation for the prediction model of brain age for brain science research on brain development and aging, and provides ideas and directions for the study of the application of brain age estimation.