AN IMPLEMENTATION OF A SPATIOTEMPORAL DEEP LEARNING ARCHITECTURE FOR DATA DRIVEN LEARNING OF BRAINS NETWORK CONNECTIVITY

Abstract

Brain disorders are often linked to disruptions in the dynamics of the brain's intrinsic functional networks. It is crucial to identify these networks and determine disruptions in their interactions to classify, understand, and possibly cure brain disorders. Brain's network interactions are commonly assessed via functional (network)\ connectivity, captured as an undirected matrix of Pearson correlation coefficients. Functional connectivity can represent static and dynamic relations. However, often these are modeled using a fixed choice for the data window. Alternatively, deep learning models may flexibly learn various representations from the same data based on the model architecture and the training task. The representations produced by deep learning models are often difficult to interpret and require additional posthoc methods, e.g., saliency maps. Also, deep learning models typically require many input samples to learn features and perform the downstream task well. This dissertation introduces deep learning architectures that work on functional MRI data to estimate disorder-specific brain network connectivity and provide high classification accuracy in discriminating controls and patients. To handle the relatively low number of labeled subjects in the field of neuroimaging, this research proposes deep learning architectures that leverage self-supervised pre-training to increase downstream classification. To increase the interpretability and avoid using a posthoc method, deep learning architectures are proposed that expose a directed graph layer representing the model's learning about relevant brain connectivity. The proposed models estimate task-specific directed connectivity matrices for each subject using the same data but training different models on their discriminative tasks. The proposed architectures are tested with multiple neuroimaging datasets to discriminate controls and patients with schizophrenia, autism, and dementia, as well as age and gender prediction. The proposed approach reveals that differences in connectivity among sensorimotor networks relative to default-mode networks are an essential indicator of dementia and gender. Dysconnectivity between networks, especially sensorimotor and visual, is linked with schizophrenic patients. However, schizophrenic patients show increased intra-network default-mode connectivity compared to healthy controls. Sensorimotor connectivity is vital for both dementia and schizophrenia prediction, but the differences are in inter and intra-network connectivity.

CHAPTER 1

INTRODUCTION

Brain disorders are often driven by disruptions in the dynamics of the brain’s intrinsic functional networks, making it extremely important to identify these networks and determine disruptions in their dynamics. For example, (Culbreth et al. 2021; Yu et al. 2011; Zhang et al. 2019; Zhu et al. 2020; Morgan et al. 2020a; Lynall et al. 2010; van den Heuvel et al. 2010) show that schizophrenic patients have high levels of functional disconnectivity between brain networks. Dysregulated brain dynamics and dysregulated dynamic connectivity across the brain’s multiple functional networks are seen in Schizophrenic patients (Supekar et al. 2019). In Alzheimer’s disease (AD), disrupted brain dynamics demonstrate cognitive dysfunction (Haan et al. 2011). (Cordova-Palomera et al. 2017) suggests that the brains of AD patients display altered oscillatory patterns and functional coupling alterations along with decreased global metastability. Alterations in brain activity have been linked to autism spectral disorder (ASD), (Just et al. 2012; Yahata et al. 2016) shows dysfunctional brain activity among the brain’s functional network for ASD patients. (Zeng et al. 2017) show significantly lower whole-brain activity for the ASD group.

It is possible to indirectly measure brain function activity to various degrees of precision using brain imaging methods, such as functional magnetic resonance imagining (fMRI). fMRI captures the nuances of spatiotemporal dynamics that could potentially provide clues to the causes of mental disorders and enable early diagnosis. However, the obtained data for a single subject is of high dimensionality (often in thousands) m and to be useful for learning, and statistical analysis, one needs to collect datasets with a large number of subjects n. Yet, for any kind of disorder, demographics, or other types of conditions, a single study is rarely able to amass datasets large enough to go out of the m ≫ n mode.

fMRI captures voxel-level data and does not provide intrinsic functional networks or their connectivity. It is extremely challenging for any machine learning (ML) or deep learning (DL) model to work directly on the voxel-level data to even perform classification between patients and healthy controls (HC). Therefore, to reduce the number of features, methods are used to get regions made up of several voxels and the connectivity between these regions. To spatially split the brain into networks, existing studies either divide the brain into multiple regions using existing pre-defined brain atlases such as Shaefer (Schaefer et al. 2017), and many others, or estimate constituent components using inference methods, such as independent component analysis (ICA) (Hyva¨rinen & Oja 2000). Whereas, the connectivity is often assessed via the functional (network) connectivity (F(N)C). Although any statistical dependence measure can be used to represent the FC or FNC, almost always FC or FNC is represented as an undirected correlation matrix of Pearson correlation coefficient (PCC) between the regions/components.

These hand-crafted features (correlation matrices) are used in studies of the brain and have demonstrated the overarching value of inspecting the brain and its disorders through the undirected weighted graph of the fMRI correlation matrix. (Yan et al. 2017) uses FC as a feature to predict schizophrenia-related changes. Whereas, (Parisot et al. 2018) use FC alongside phenotypic and imaging data as inputs to extract graph features for the classification of AD and Autism. (Kawahara et al. 2016) uses connection strength between brain regions as edges, typically defined as the number of white-matter tracts connecting the regions. (Ktena et al. 2017) employs spectral graph theory to learn similarity metrics among functional connectivity networks.

ML and DL methods can use FC matrices to perform classification between patients and HC with high accuracy. Many studies use FC to predict the gender or disease/disorder (Arslan et al. 2018; Kazi et al. 2021; Kim & Ye 2020; Ktena et al. 2018; Ma et al. 2019) using graph neural networks (GNNs) or other such methods. However, the dynamics of brain function vanishes into proxy features such as correlation matrices of FC. Correlation-based FC matrices have many shortcomings including but not limited to inflexibility in terms of the downstream task, undirected relations among regions and networks, and limitation in capturing temporal dynamics.

One of the aims of this research is to show that dynamic DL architectures can be created that work directly on the BOLD time courses and learn task-dependent directed connectivity structures between networks for individual subjects. Interpretation of these estimated connectivity structures could lead to useful insights regarding brain functionality and multiple brain disorders.

This dissertation presents these DL architectures and shows that DL architectures without using hand-crafted features can beat SOTA ML and DL methods that use hand-crafted features in discriminating controls and patients with schizophrenia, autism, and dementia, as well as age and gender prediction from functional MRI data. More importantly, this work shows that connectivity matrices estimated by our DL architectures are more interpretable, are robust to confounding factors, show the direction of connectivity between networks, and capture more temporal dynamic states than correlation-based FC matrices.

Statement of the Problem

 

Despite significant advancements in brain imaging technologies, understanding the dynamics of intrinsic functional networks and their disruptions in brain disorders remains a challenging task. Brain disorders, such as schizophrenia, Alzheimer's disease, and autism spectrum disorder, are often driven by alterations in the brain's functional connectivity, leading to cognitive and behavioral abnormalities. Identifying and characterizing these disruptions is crucial for early diagnosis, personalized treatment, and gaining insights into the underlying mechanisms of these disorders.

 

Current research relies heavily on hand-crafted features, such as correlation matrices of functional connectivity, to study brain disorders. While these approaches have demonstrated their value in discriminating between patients and healthy controls, they have limitations in terms of flexibility, interpretability, and capturing temporal dynamics. Additionally, due to the high dimensionality of brain imaging data and the need for large datasets, single studies often struggle to collect sufficient samples, hindering the development of accurate and robust diagnostic models.

 

This research addresses the following critical challenges:

1. The need to identify and understand the disruptions in intrinsic functional networks of the brain in patients with various disorders, such as schizophrenia, Alzheimer's disease, and autism spectrum disorder.
2. The limitations of using hand-crafted features, such as correlation matrices, in machine learning and deep learning models for classification and prediction tasks related to brain disorders.
3. The lack of dynamic deep learning architectures that can work directly on BOLD time courses and learn task-dependent directed connectivity structures between networks for individual subjects.
4. The requirement for more interpretable and robust models that can capture the temporal dynamics of brain function and effectively distinguish between controls and patients with different brain disorders.

By addressing these challenges, this research aims to pave the way for more effective and accurate diagnostic tools and contribute to a better understanding of brain functionality and the mechanisms underlying various brain disorders.

Research Aims:

1. To design and develop dynamic deep learning architectures that can work directly on the BOLD time courses obtained from functional magnetic resonance imaging (fMRI) data. These architectures should be capable of learning task-dependent directed connectivity structures between intrinsic functional networks for individual subjects. By creating these novel architectures, we seek to enhance the interpretability and capture more temporal dynamics of brain function compared to existing hand-crafted feature-based methods.

 

2. Improve Classification and Prediction for Brain Disorders: Another key aim of this research is to leverage the dynamic deep learning architectures to improve classification and prediction tasks related to brain disorders. We aim to demonstrate that these architectures, without relying on hand-crafted features like correlation matrices, can outperform state-of-the-art machine learning and deep learning models in discriminating between patients and healthy controls for schizophrenia, Alzheimer's disease, and autism spectrum disorder. Additionally, we will explore the use of these architectures for age and gender prediction from fMRI data.

 

3. Enhance Interpretability and Robustness: The third research aim is to achieve better interpretability and robustness in the estimated connectivity structures. We will investigate how dynamic deep learning architectures can provide more interpretable insights into brain functionality and the disruptions observed in various brain disorders. Moreover, we will assess the robustness of these connectivity matrices to confounding factors and their ability to accurately represent the direction of connectivity between different brain networks. By addressing these objectives, we aim to contribute to a deeper understanding of brain disorders and the underlying mechanisms driving them.

 

AN IMPLEMENTATION OF A SPATIOTEMPORAL DEEP LEARNING ARCHITECTURE FOR DATA-DRIVEN LEARNING OF BRAINS NETWORK CONNECTIVITY. GET MORE COMPUTER SCIENCE PROJECT TOPICS AND MATERIALS

Sharing is caring!