毕业生知识库

功能磁共振成像(functional Magnetic Resonance Imaging, fMRI)作为一种前沿的脑功能成像技术，克服了传统上仅依赖神经解剖学理解大脑功能的局限，能够揭示大脑的时空模式，为精神疾病的客观诊断与治疗提供了独特的机会。然而，如何有效挖掘并利用此类高维度、复杂的神经影像数据成为一项挑战。近年来，深度学习作为新兴的人工智能技术，以其自动从高维数据中学习最优特征的能力，大大简化了传统机器学习中的特征工程步骤，为处理和分析fMRI数据提供了一种自动化且高效的解决方案，同时也为神经影像学研究带来了新的挑战。因此，本课题旨在开发适用于功能磁共振影像的深度学习算法，自动挖掘影像数据中隐藏的脑功能活动信息，为精神疾病的客观诊断提供创新性的工具和方法，并力求发现客观、有效的影像学标志物，以助于深化对精神疾病发病机理的理解。论文的研究内容主要围绕三个核心创新点展开，旨在应对当前精神疾病诊断中面临的fMRI动态信息丢失、单一特征局限及多特征融合效率低下等问题：

(1) 研究内容一：本研究提出了基于时间序列的融合分类与聚类的深度学习框架。已有基于深度学习的精神疾病诊断方法主要依赖于功能连接分析，其计算相关性的过程可能导致原始动态信息的丢失。而且多数研究仅关注单疾病与健康对照的二分类问题，未能充分探究不同精神疾病潜在的内部联系。基于此，本研究采用可以直接建模时间序列的深度学习模型MsRNN，并创新性地将分类与聚类集成在一个框架下，利用MsRNN学到的具有分辨力的特征进一步聚类分析，深入挖掘不同精神疾病之间的内在联系。在包含693名被试的BSNIP数据集上，本研究通过与现有方法的分类性能对比验证了MsRNN模型的有效性。此外，借助留一特征法，我们识别出了对分类结果有决定性影响的关键脑区组成部分。在此基础上，本研究对聚类得出的疾病亚型针对临床指标(如症状评分)进行差异性分析，尝试建立亚型聚类结果与临床实际疾病分型之间的映射关系，为精神疾病的诊疗提供了更为细致的认识。结果发现分裂情感障碍可以分为两个截然不同的亚型，其中一个亚型在情绪退缩、焦虑、内疚感等临床症状评分显著高于另一个亚型，且表现出更多双极症状特点。

(2) 研究内容二：本研究发展了可以融合脑功能连接和时间序列的深度学习框架。尽管MsRNN等基于深度学习的精神疾病诊断方法已取得一定的成果，但是普遍局限于单一特征，忽视了不同特征间潜在的互补性。鉴于此，本开发能够融合脑功能连接和时间序列信息的深度学习框架HDLFCA，利用卷积-循环神经网络和深度神经网络分别学习fMRI数据中的时间动态信息和功能耦合信息，然后进一步通过逻辑回归分类器整合这两种信息提升脑疾病分类准确率。为进一步提升模型的可解释性，本研究引入注意力机制，在训练过程中自适应地学习重要脑区，实现疾病诊断与关键特征定位的一体化，既提高了分类性能，又找出了对分类有重大贡献的脑区特征。在包含1100名被试的私有精神分裂症数据集和1522名被试的公开ABIDE数据集上的验证实验表明，HDLFCA超越了12种主流分类方法，准确率指标提高了2.8%至8.9%，验证了其有效性。此外，本研究还发现模型预测的患病概率与临床症状间存在显著相关性，有力地证明了HDLFCA框架在解析精神疾病发病机制方面的价值。

(3) 研究内容三：基于互学习本研究构建了可以融合多特征的端到端深度学习模型。尽管目前存在少量基于深度学习的精神疾病诊断研究尝试利用多特征进行分类(例如研究内容二的HDLFCA)，但大多采取两阶段策略，如先独立训练多个特征对应的分类器再通过集成学习方法综合结果。这种分阶段的方法往往耗时较长，且对各个特征之间的关联性没有充分利用。基于此，本研究基于互学习开发可以融合脑功能连接和时间序列的端到端深度学习模型。该架构使处理不同特征的网络能够协同学习，挖掘两者间共享的、鲁棒的特征表达，有效提升了时间效率的同时保持了优异的分类性能。此外，为了更好地捕捉脑区间的全局相关性，本研究还引入了Transformer这种能够有效捕捉序列数据中长距离依赖关系的模型。实验上，本研究在两个多站点fMRI数据集上与16种基线模型进行对比，结果显示CFML显著优于所有基线方法，准确率提高了3.0%-9.2%。相较于两阶段方法HDLFCA，CFML的训练速度快4倍的同时还保持了卓越的分类性能，验证了方法的有效性。通过注意力图的可视化，本研究揭示了最具辨别力的脑区特征，这些特征与既往研究相符，进一步验证了CFML在识别潜在影像标记物方面的潜能。

综上所述，本论文通过系统地开展上述三项研究，逐一解决了精神疾病诊断中动态信息损失、单一特征依赖和多特征融合低效等关键问题，不仅创新性地开发了新的深度学习模型以提升分类精度，还强调了模型的可解释性和对疾病内在机制的深入理解，从而为精神疾病的精准诊断提供了有力的工具和崭新的视角。

Functional Magnetic Resonance Imaging (fMRI) stands at the vanguard of brain imaging techniques, transcending traditional neuroanatomical analyses to elucidate brain functionalities. By dissecting fMRI data, this approach sheds light on crucial temporal and spatial patterns, thereby providing invaluable insights for the objective diagnosis and treatment of mental disorders. Despite its potential, navigating the labyrinth of high-dimensional, complex neuroimaging data poses a significant challenge. Recent strides in deep learning can autonomously extract optimal features from large datasets and have refined conventional machine learning methodologies, offering a more efficient, automated, and robust framework for analyzing fMRI data. This progress paves the way for new avenues in neuroimaging research and introduces fresh challenges. This dissertation delves into the development of deep learning algorithms tailored for MRI data, aiming to uncover hidden patterns of brain activity. It pioneers tools and methodologies for objectively diagnosing mental disorders, potentially identifying reliable and effective imaging biomarkers that enhance our comprehension of disease pathophysiology. The thesis systematically tackles prevalent diagnostic issues in mental health, such as the reduction of original signal dynamics, an overdependence on single features, and the challenges in integrating multiple features:

1. We introduce a novel deep learning framework that integrates classification and clustering, targeting time series data. Contrary to traditional methods that often reduce time series data to mere correlation measures—thereby omitting raw dynamic content—this segment proposes MsRNN, an innovative model adept at directly modeling time series for complex mental illness diagnostics. Demonstrated through empirical validation on the BSNIP dataset (693 subjects) and comparative analyses, MsRNN not only excels in classification accuracy but also innovatively combines classification and clustering. Through a leave-one-feature-out strategy, it pinpoints critical brain regions impacting classification outcomes, offering deeper insights into mental disease correlations and refining diagnostics and treatment approaches through clinical indicators and symptom score analysis.

2. We present a groundbreaking deep learning framework that fuses brain functional connectivity with time series data. Moving beyond the advances of MsRNN, which predominantly concentrated on isolated features, this chapter reveals the HDLFCA framework. This first-of-its-kind framework integrates fMRI dynamics and functional connectivity using Convolutional-Recurrent Neural Networks and Deep Neural Networks for learning temporal and functional information, respectively. It not only improves disease classification performance but also employs attention mechanisms for identifying significant brain regions during diagnosis, enhancing accuracy and underscoring influential classification features. Its effectiveness is validated across two extensive datasets, surpassing 11 benchmark models and demonstrating a significant correlation between disease probabilities and clinical symptoms.

3. We develop an unprecedented end-to-end deep learning model for multi-feature fusion based on mutual learning. This model addresses the time-consuming and fragmented nature of previous feature-fusion methods by fostering collaborative feature representation extraction across different networks. With the introduction of the Transformer model, it captures global inter-regional dependencies, highlighting the model's superior multi-feature utilization and efficiency over traditional two-stage methods. Through explainable algorithms, this study unveils key features impacting classification.

In summary, this thesis methodically addresses critical challenges in diagnosing mental disorders, such as the loss of dynamic information, reliance on singular features, and inefficiencies in multi-feature fusion. It pioneers innovative deep learning models that not only improve classification accuracy but also enhance interpretability and deepen our understanding of disease mechanisms. These contributions provide powerful tools and new perspectives for accurate diagnostics of mental health disorders.