Modern neuroimaging analysis is an important research direction for understanding the information-processing mechanism of human brain in a non-invasive manner. It has become one of the most powerful tools for diagnosing complex mental disorders such as depression, autism spectrum disorder (ASD), and attention deficit hyperactivity disorder (ADHD). However, those disorders exhibit diverse and complex symptoms that are challenging to diagnose and treat using conventional clinical manifestations. Recent advances in neuroimaging techniques, such as electroencephalography (EEG), functional magnetic resonance imaging (fMRI), and functional near-infrared spectroscopy (fNIRS), have sprung up to various deep representation learning methods that effectively and automatically diagnose mental disorders.

Despite the advantages, the high cost and time-consuming aspect of gathering neuroimaging data often limits sample sizes and introduces uncontrolled confounding factors in existing neuroimaging studies. As a result, such studies can suffer from limited statistical power and reduced generalizability, compromising their diagnostic robustness in real-world scenarios. This thesis focuses on developing effective and robust neuroimaging analysis models by comprehensively representing neuroimaging data from multiple perspectives. Specifically, the proposed models address the above challenges by increasing training samples (through multi-task learning) and enhancing feature representation capabilities (using multi-expert, multi-view, and multi-modal methods).

Firstly, this thesis presents a comprehensive analysis of diverse neuroimaging analysis techniques that employ deep learning or machine learning approaches, including multi-task learning, dynamic learning, interpretable learning, and graph representation learning. The challenges and limitations in current deep neuroimaging analysis, particularly in its clinical application, are elucidated.

Secondly, a series of effective deep learning models are proposed from different perspectives to enhance the efficiency of mental disorder diagnosis based on MRI data. It is worth noting that this thesis highlights medical interpretability in developing these models to ensure their practical application in clinical settings. Specifically, a multi-task learning framework is proposed for the joint diagnosis of ASD and ADHD, exploiting the significantly overlapping symptoms between these two disorders. By exploiting multi-task learning paradigm, we train our model on datasets from two public diseases simultaneously, thereby increasing the sample size for neuroimaging analysis and learning more discriminative representations through induced coupling. For precise diagnosis of ADHD, we present a multi-expert network with brain region clustering to comprehensively captures the complex feature patterns of neuroimaging data. Our simulation experiments demonstrate that employing different expert networks to capture the unique characteristics of different brain regions can significantly enhance the diagnostic accuracy of ADHD. To account for the spatio-temporal characteristics of neuroimaging data, we design a spatio-temporal co-attention learning framework that can simultaneously model the spatio-temporal correlation of fMRI data for the diagnosis of ASD and ADHD from both spatial and temporal views. This framework can automatically locate remarkable regions of interest and time frames in time-series fMRI data, facilitating the exploration of neurological patterns associated with specific mental disorders through the co-attention mechanism. Building on our spatio-temporal framework, a multi-modal approach is proposed to employ graph neural networks and Transformers to capture the discriminative spatial and temporal representations of neuroimaging data. This multi-modal approach can integrate the advantages of different modalities to further enhance the diagnostic accuracy of mental disorders.

Lastly, while the proposed models demonstrate reliable performance in diagnosing ASD and ADHD, there is currently not any universally general approach to efficiently analyze diverse neuroimaging data, which can reduce the training cost. Therefore, a cross-platform and cross-task neuroimaging analysis platform is presented to effectively learn discriminative dynamic representations of neuroimaging data based on graph representation learning.

Overall, the proposed models have been demonstrated to be effective and robust through extensive experimental results on several real-world datasets. Additionally, they also enable the interpretable neuroimaging analysis to explore salient disorder-specific neuroimaging patterns. These findings provide valuable insights for neuroimaging professionals and contribute to advancing neuroimaging analysis in clinical applications.