Recently, the human connectome (the structural and functional connectivity patterns of the human brain) has aroused great interests of neuroscientists and attracted large amounts of investment and public attention (e.g. the Brain Initiative Project). Currently, the analysis of non-invasive neuroimaging data has faced two critical challenges. Firstly, an increasing number and size of datasets are generated in the community, which leads to very high requirements for the computational capabilities. Second, the high-resolution neuroimaging data requires more efficient and reasonable statistical methods for the high dimensional data analysis.
In response to these transformations, our research interests start from three aspects: a heterogeneous platform with high computational capabilities, efficient methods for high-dimensional neuroimaging data analysis, and emerging applications in cognitive neuroscience and artificial intelligence.
A Hybrid CPU-GPU Platform for Fast Mapping of High-resolution Human Connectome
An important and promising way to study the human connectome is to combine non-invasive neuroimaging techniques and graph theoretical approaches. However, an increasing number and size of datasets are generated in the community currently, and there are very high requirements for the computational capabilities in the study of human connectome, especially for the high-resolution neuroimaging data analysis. We propose a hybrid CPU-GPU framework to accelerate the mapping of the human brain connectome. We applied this framework to a publicly available resting-state functional MRI dataset. The results demonstrated that our hybrid framework saved a tremendous amount of time for such a graph theoretical network analysis method. Further analyses revealed some topological properties of high-resolution functional brain networks that are largely compatible with previous ROI-based human brain network studies. We also present a GPU-based accelerating framework for brain fiber tractography based on diffusion tensor MRI. In the probabilistic streamlining fiber tracking, we find that fiber lengths are exponentially distributed, and propose a novel segmenting strategy to improve the load balance. On mid-range GPUs, we achieve performance gains up to 30x over CPUs for the whole process. Taken together, our proposed framework can substantially enhance the applicability and efficacy of high-resolution (voxel-based) brain network analysis, and have the potential to accelerate the mapping of the human brain connectome in normal and disease states.
Graphical Model for Voxel-wise Brain Network and Feature Extraction
The brain is preferable to be modeled as a complex network (or graph) that contains a large quantity of nodes and connections. The brain nodes are usually defined by imaging voxels or regions of interest (ROIs); the brain edges are defined by measuring the structural or functional association between the nodes based on neuroimaging data. Voxel-based brain networks possess more naturally defined nodes (i.e., imaging voxels) in a higher spatial resolution, which may unveil more detailed connectivity information especially for regions that contain multiple sub-divisions and reduce the effects of the possible spatial inhomogeneity in ROI-based studies. However, to construct a brain network using graphical model, we need to face the curse of dimensionality, especially for the voxel-wise functional network. Currently, we are looking for more efficient and precise graphical estimation and inference methods for functional brain networks instead of the traditional approach using Pearson correlation matrix. We are also interested in the feature extraction from brain networks in both normal and disease states. For example, the functional hub distribution may be a critical feature for the cognitive learning of human brain (Fig. 2), and can be disturbed by neuropsychiatric disorders. Our ultimate goal is to learn more about the brain features of different subjects and find useful and robust biomarkers for multiple diseases.
Emerging Applications in Cognitive Neuroscience and Artificial Intelligence
We intend to validate the efficiency of our platform and methods in several emerging applications in cognitive neuroscience and artificial intelligence.
First of all, reliable measures with both low intra-subject and high inter-subject variability are essential for both clinical and theoretical study. Although the studies of human connectome using graph theory have provided insights into the topological organization of the human brain in health and disease, the test-retest reliability of the graph metrics in brain networks remains to be further investigated, especially in voxel-based whole-brain functional networks during resting-state. We studied the short- and long-term test-retest reliability of high-resolution functional connectomics using a resting-state fMRI dataset containing 53 young healthy adults. The test-retest reliabilities of network metrics were sensitive to the scanning orders and intervals. Most network metrics were generally test-retest reliable in long-term case, with the highest reliability in the clustering coefficient among global metrics and in nodal degree and efficiency among the nodal metrics. Our study provides important guidance for choosing reliable network metrics and analysis strategies in future studies.
In future studies, our research will focus on categorizing brain image data from subjects in healthy state and multiple disorders, using state-of-the-art statistical machine learning methodologies. We are also interested in the topological structure of brain network, and the role it plays in cognitive and learning task. The study can provide potential insights into brain-inspired computation, which is a major research direction of our lab.
- HDC-IM: Hyperdimensional Computing In-Memory Architecture based on RRAM , 2019. pdf slide
- PAGANI Toolkit: Parallel graph-theoretical analysis package for brain network big data , in Human brain mapping, vol.1, No.17, 2018. pdf
- Balanced quantization: An effective and efficient approach to quantized neural networks , in Journal of Computer Science and Technology, vol.32, No.4, 2017, pp.667-682. pdf
- A GPU-accelerated Framework for Fast Mapping of Dense Functional Connectomes , in The Second International Conference on Neuroscience and Cognitive Brain Information, accepted as a full paper, 2017. pdf
- Line propagation based on FDT probabilistic tracking (LPFPT) , in 2016 Annual Meeting of the Organization for Human Brain Mapping, accepted as a poster, 2016. pdf slide
- PAGANI Toolkit: Parallel Computing Package for Fast Network Analyses of Brain Connectomes , in 2016 Annual Meeting of the Organization for Human Brain Mapping, accepted as a poster, 2016. pdf slide
- Test-Retest Reliability of Graph Metrics in High-Resolution Functional Connectomics: A Resting-State Functional MRI Study , in CNS Neuroscience & Therapeutics, vol.21, No.10, 2015, pp.802-816. pdf
- A Hybrid CPU-GPU Accelerated Framework for Fast Mapping of High-Resolution Human Brain Connectome , in PloS one, vol.8, No.9, 2013, pp.e62789. pdf
- A reconfigurable computing approach for efficient and scalable parallel graph exploration , in Proceedings of the IEEE 23rd International Conference on Application-Specific Systems, Architectures and Processors (ASAP), 2012, pp.8-15. pdf
- Parallel FPGA-based all pairs shortest paths for sparse networks: A human brain connectome case study , in Proceedings of 22nd International Conference on Field Programmable Logic and Applications (FPL), 2012, pp.99-104. pdf
- Probabilistic brain fiber tractography on gpus , in Proceedings of the IEEE 26th International Parallel and Distributed Processing Symposium Workshops PhD Forum (IPDPSW), 2012, pp.742-751. pdf
- Node importance analysis in complex networks based on hardware computing , in Dianzi Yu Xinxi Xuebao (Journal of Electronics and Information Technology), vol.33, No.10, 2011, pp.2536-2540.
- A heterogeneous accelerator platform for multi-subject voxel-based brain network analysis , in Proceedings of the International Conference on Computer-Aided Design (ICCAD), 2011, pp.339-344. pdf
- Hardware computing for brain network analysis , in Proceedings of the 2nd Asia Symposium on Quality Electronic Design (ASQED), 2010, pp.219-222. pdf