脑网络组研究知识库

作为一种无创的成像技术，功能磁共振影像(functional magnetic resonance imaging, fMRI)通过检测血氧浓度来推断大脑的活动状态，在认知科学以及精神疾病研究中被广泛应用。脑功能磁共振影像具有高维小样本、低信噪比等特点，且精神疾病具有脑损伤呈弥散性和症状重叠等特性，因此为基于MRI的精神疾病自动判别分类模型的性能提升带来了诸多挑战。近年来，深度学习技术高速发展，为进一步解决精神疾病的分类诊断提供了新方法。本课题致力于开发适用于fMRI的深度学习技术，以实现对MRI影像的深入量化分析，为辅助临床实现精神疾病的早期分类判别，发掘客观、有效的影像学标志物作出贡献。本文的主要创新性工作如下：

1. 提出了适用于静态脑功能网络连接的L_1,2范数正则化深度神经网络(deep neural network, DNN)，在精神分裂症的多中心分类中实现跨中心预测平均精度超过81%。并首次使用逐层关联传递法(layer-wise relevance propagation, LRP)实现了对疾病相关脑区的追溯。

L_1,2范数正则化深度神经网络模型以被试的功能网络连接特征作为输入，结合L_1,2范数正则化、dropout、batch normalization等模型训练策略，在1100例被试的汉族多中心大样本数据集上实现了对精神分裂症(schizophrenia，SZ)的精确诊断。相比于支持向量机、随机森林和Adaboost等经典的机器学习算法，DNN对SZ的诊断性能有了显著的提升。将LRP算法应用于深度学习模型的解释，追溯与疾病相关的特异性脑功能网络，包括额叶网络、皮下网络以及小脑网络等。

2. 提出了适用于动态脑功能连接(dynamic functional connectivity, dFC)的双向全拼接长短时记忆模型(Full-BiLSTM)，实现了更高精度的轻度认知障碍(mild cognitive impairment, MCI)早期识别率(73%)。

相较于使用静态功能连接，dFC蕴含了更丰富的时变信息。此研究首先使用基于滑动窗的方法计算dFC，将其作为Full-BiLSTM的输入特征进行训练和测试。此模型首次实现了深度学习对于动态功能连接的分类任务。将此模型应用于Alzheimer's Disease Neuroimaging Initiative数据集的MCI分类任务，准确率达到73%。实验证明了Full-BiLSTM可以有效地从复杂的时间序列中捕捉与特征相关的脑动态变化。

3. 提出了适用于fMRI脑功能时间序列(time course, TC)的多尺度卷积循环神经网络MsRNN，并利用“去一特征法”对深度学习模型进行解释，实现了对疾病判别分类最具贡献性脑区的追溯。

相较于静/动态功能连接，TC序列包含了磁共振影像记录到的更为完整的脑活动信息。此研究首先从原始fMRI中提取每个被试的TC，将其作为MsRNN的输入特征进行训练和测试。MsRNN的卷积层自动学习脑区之间的空间关联，省略了经典脑网络分析中计算功能连接的步骤，避免了在功能连接计算时导致的信息丢失问题。循环网络层自动整合时域信息，从而实现了对TC“时间-空间”特征的联合分析。MsRNN在1100例被试的汉族多中心大样本数据集上实现了对SZ的精确分类(准确率83%)。使用提出的“去一特征法”精确追溯对SZ分类贡献度最高的脑区主要集中于背侧纹状体、壳核、小脑等。

综上所述，本课题创造性地提出了三种适用于fMRI的深度学习算法，并结合不同fMRI影像特征设计改进了算法的可解释性。在实际多种精神疾病的多中心分类任务中，本文提出的算法模型均优于已有的多种经典分类方法，有望发现可用的影像学标志物，为精神疾病的早期分类判别和正确诊断提供方法学支撑。

As a non-invasive imaging technology, functional magnetic resonance (fMRI) has been widely used in cognitive science and mental disease research to measure brain activity by detecting blood oxygen level-dependent. Brain functional magnetic resonance imaging has the characteristics of high dimensional, small samples, low signal-to-noise ratio, and mental illness has the characteristics of brain injury such as diffusion and symptom overlap, which brings many challenges to the performance improvement of the automatic discrimination classification model of mental illness based on MRI. In recent years, the development of deep learning technology provides a new way to solve the classification and diagnosis of mental illness. The project is to develop new deep learning methods for fMRI, realize the quantitative analysis of MRI, and contribute to the early classification and discrimination of mental diseases. The main innovative points of this paper are as follow:

1. The L_1,2-norm regularized deep neural network model is proposed, which takes the functional network connection characteristics of the subjects as the input, and combines the l1,2-norm regularization, dropout, batch normalization and other model training strategies to realize the accurate diagnosis of schizophrenia (SZ) on the multi center large sample dataset of 1100 Han subjects. Compared with the classical machine learning algorithms such as Support Vector Machine, random forest and AdaBoost, DNN has a significant improvement on the diagnosis performance of SZ. LRP algorithm is applied to the interpretation of deep learning model to trace the specific brain function connections related to diseases, including frontal network, subortical network and cerebellar network.

2. A Fully Concatenated Bidirectional LSTM (Full-BiLSTM) for dynamic functional connectivity (dFC) is proposed to for identifying early MCI with the accuracy of 73%.

Compared with static function connection, dFC contains more time-varying information. In this study, we first use the sliding window method to calculate the dFC, which is then used as the input features of Full-BiLSTM for training and testing. The model is applied to the diagnosis of MCI patients. Experiments show that Full-BiLSTM can effectively capture the brain dynamic changes related to features from complex time series, and has obvious advantages in using the time-varying information contained in fMRI.

3. A multi-scale convolutional cyclic neural network msrnn for time course (TC) of fMRI is proposed, and the "leave-one-feature-out method" is used to trace the most contributing brain area of disease discrimination and classification, which improves the interpretability of deep learning model.

Compared with static/dynamic functional connection, TC sequence contains more complete brain activity information recorded by MRI. In this study, the TCs of each subject are extracted from the original fMRI, which is used as the input feature of MsRNN for training and testing. The convolution layer of msrnn can automatically learn the spatial association between brain regions, which goes beyond the steps of calculating functional connection in classic brain network analysis, and avoids the problem of information loss caused by the calculation of functional connection. The cyclic network layer of the model automatically integrates the time-domain information, thus realizing the spatio-temporal analysis of TC. MsRNN achieved the classification accuracy of 83% on 1100 multi center large sample datasets of Han People. Using the proposed "leave-one-feature-out" to trace the brain regions that contribute most to SZ classification, the regions mainly located at dorsal striatum, putamen and cerebellum.

In conclusion, three deep learning algorithms for different fMRI features are proposed, In the multi-center classification task, the algorithm model proposed in this paper is superior to the existing classical classification methods, and it is expected to find the available imaging markers.