Empower Deep Learning for Brain Network Analysis Restricted; Files Only

Abstract

Recent large-scale brain network datasets, such as the Philadelphia Neurodevelopmental Cohort (PNC) study and the Adolescent Brain Cognitive Development (ABCD) study, have laid the foundation for applying deep learning techniques to brain network analysis. These datasets provide extensive and diverse brain imaging data and rich phenotypic information, enabling researchers to investigate the complex relationships between brain networks and behavioral measures in large populations. However, applying deep learning to brain network analysis poses several challenges, including the need for better backbone architectures, sample size limitations, and limited supervision signals. This thesis aims to address these challenges by developing innovative deep-learning techniques spanning both model architectures and training strategies.

In the first part of the thesis, we focus on designing novel model architectures tailored for brain network analysis. We propose FBNetGen, an end-to-end differentiable pipeline that generates task-aware functional brain networks from raw fMRI time series data, achieving good performance and providing explainable insights into disorder-specific brain regions and connections. We then introduce Brain Network Transformer (BNT), a transformer-based architecture designed to capture the unique properties of brain networks, demonstrating superior performance on large-scale fMRI datasets. Furthermore, we present Dynamic bRAin Transformer (DRAT), an approach that focuses on modeling dynamic brain networks to capture temporal variations and improve predictions and interpretability.

The second part of the thesis focuses on advanced training strategies to enhance the generalization and performance of deep learning models for brain network analysis. We develop R-mixup, a data augmentation approach operating on the Riemannian manifold of symmetric positive definite matrices, effectively addressing the limited sample size challenge in low-resource settings commonly encountered in neuroimaging studies. Additionally, to obtain richer supervision signals, we propose a multi-task learning framework that jointly predicts various behavioral and clinical measures from brain networks, enabling knowledge sharing across related tasks and improving individual task performance while better utilizing the wide variety of annotated measures available in existing datasets.

Extensive experiments on multiple datasets and tasks demonstrate the superior performance and practical value of our methods. This thesis's contributions facilitate a better understanding of the complex relationships between brain networks and behavioral phenotypes, benefiting neuroimaging research and clinical applications. By addressing the key challenges of better backbone architectures, sample size limitations, and limited supervision signals, this thesis paves the way for more effective and explainable deep learning techniques in brain network analysis, ultimately advancing our understanding of the human brain and its role in cognition and disorders.

Table of Contents

1 Introduction

1.1 The Importance of Brain Network Analysis

1.2 The Rise of Deep Learning in Brain Network Analysis

1.3 Large-Scale Brain Network Datasets

1.4 Challenges and Solutions in Applying Deep Learning to Brain Network Analysis

1.4.1 Better Backbone Architectures

1.4.2 Sample Size Limitations

1.4.3 Limited Supervision Signals

2 Model Architecture: Task-aware GNN-based fMRI Analysis via Functional Brain NETwork GENeration (FBNetGen)

2.1 Introduction

2.2 Background and Related Work

2.2.1 fMRI-based Brain Network Analysis

2.2.2 Graph Neural Networks

2.3 FBNetGen

2.3.1 Overview

2.3.2 Feature Encoder

2.3.3 Graph Generator

2.3.4 Graph Predictor

2.3.5 End-to-end Training

2.4 Experiments

2.4.1 Experimental Settings

2.4.2 RQ1: Performance Comparison

2.4.3 RQ2: Ablation Studies

2.4.4 RQ3: Influence of Hyper-parameters

2.4.5 Interpretability Analysis

2.5 Conclusion

3 Model Architecture: Brain Network Transformer

3.1 Introduction

3.2 Background and Related Work

3.2.1 GNNs for Brain Network Analysis

3.2.2 Graph Transformer

3.3 Brain Network Transformer

3.3.1 Problem Definition

3.3.2 Multi-Head Self-Attention Module (MHSA)

3.3.3 Orthonormal Clustering Readout (OCRead)

3.3.4 Generalizing OCRead to Other Graph Tasks and Domains

3.4 Experiments

3.4.1 Experimental Settings

3.4.2 RQ1: Performance Analysis

3.4.3 RQ2: Ablation Studies on the OCRead Module

3.4.4 RQ3: In-depth Analysis of Attention Scores and Cluster Assignments

3.5 Discussion and Conclusion

4 Model Architecture: DynAmic bRain Transformer (DART) with

Multi-level Attention for Functional Brain Network Analysis

4.1 Introduction

4.2 Method

4.3 Experiments

4.3.1 Experimental Settings

4.3.2 Performance and Analysis

4.3.3 Attention Visualization and Analysis

4.4 Conclusion 

5 Data Augmentation: Riemannian Mixup for Improved Generalization 

5.1 Introduction

5.2 Related Work

5.2.1 Mixup for Data Augmentation

5.2.2 Geometric Deep Learning

5.3 R-Mixup

5.3.1 Notations and Preliminary Results

5.3.2 R-Mixup Deduction

5.3.3 Comparison with Other Metrics

5.3.4 R-Mixup Theoretical Justification

5.3.5 Time Complexity and Optimization

5.4 Experiments

5.4.1 Experimental Setup

5.4.2 RQ1: Performance Comparison 

5.4.3 RQ2: The Relations of Sequence Length, SPD-ness and Model Performance

5.4.4 RQ3: Hyperparameter and Efficiency Study

5.5 Conclusion

6 Multi-Task Learning Framework: Leveraging Diverse Prediction Targets for Enhanced Individual Task Performance and Dataset Utilization

6.1 Introduction

6.2 Method

6.2.1 Problem Definition

6.2.2 Model Architecture

6.2.3 Multi-task Training Strategies

6.3 Experiments

6.4 Task Correlation Analysis

6.5 Conclusion

7 Conclusion 

7.1 Summary of Achievements

7.2 Future Directions

7.2.1 Causality and Effective Connectome for Brain Network Analysis 

7.2.2 Few-shot Learning for Extremely Unbalanced Tasks

7.2.3 Comprehensive and Clinical Evaluation

About this Dissertation

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Computer Science and Informatics
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Biology, Neuroscience Computer Science
Stichwort	Deep Learning Machine Learning fMRI Brain Network
Committee Chair / Thesis Advisor	Yang, Carl, Emory University
Committee Members	Zhao, Liang, Emory University Li, Xiaoxiao, The University of British Columbia Guo, Ying, Emory University

Zuletzt geändert

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 22 November 2024	2024-04-23 12:29:14 -0400	File download under embargo until 22 November 2024

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions