- Welcome to my GitHub profile!
- 🤔 I am a PhD student in Biomedical Engineering, working with Prof. Guoshi Li at the University of North Carolina at Chapel Hill, Chapel Hill, NC, USA.
- "Attendre et espérer.”
- 📫 Reach me at gsonwu@gmail.com
Computational Neuroscience MRI technology CV in medical imaging Optimization theory
-
In progress 09/ 2025 -
University of NorthCarolina at Chapel Hill (UNC)
- Degree: Ph.D. in Biomedical Engineering
- Location: Chapel Hill, NC, USA
-
09/2023 - 05/2025
University of Southern California (USC)
- Degree: Master of Science in Biomedical Engineering
- Location: Los Angeles, California, USA
- GPA: 3.7/4.0
-
09/2019 - 06/2023
Southern University of Science and Technology (SUSTech)
- Degree: Bachelor of Science in Biomedical Engineering
- Location: Shenzhen, Guangdong, China
- GPA: 3.5/4.0
- Thesis: Network control theory to identify the critical nodes and input signals from task fMRI
- Neural Computing and Control Laboratory. Supervisor: Prof. Quanying Liu
- Fan Magnetic Resonance (MR) Imaging Research Lab. Supervisor: Prof. Zhaoyang Fan
- Valero Lab. Supervisor: Franciso Valero-Curevas
- TRANSCEND. Supervisor: Dr. Guoshi Li
- Develope a framework for non-invasive estimation of excitatory-inhibitory balance for fMRI
- Using test-time adaptation (TTA) to predict the consistency level of the pituitary adenomas
- About: The effects of the deep learning would drop a lot when the testing set has a large different distribution with the training domain, especially when the size of the training is small, which is a common scenario in the medical image processing. TTA is a kind of technique which can improve the model performance in the datasets from other domains, without requiring the knowledge from the original domain and can be adapted to multi testing domain. In this study, we would train a based segmentation model with TTA, then use the transfer learning to change the head of the segmentation model to predict the consistency level of the pituitary adenomas.
- Method: Dataset is splitted into training (data from a single institution) and testing (data from multi institutions). The segmentation model would be trained on the training data at first, and then include the TTA during the testing. After that, the segmentation model is freezed and transfer learning is applied to predict the consistency level.
-
Use network control theory to identify the critical nodes and input signals from task fMRI.
- About: The human brain is an orderly dynamic system and it coordinates task-related regions hierarchically to perform a complex cognitive task. However, the underlying regulation mechanisms of how the brain organizes these neural circuits remain elusive from a computational perspective. Brain network control theory provides a basic theoretical architecture for linking brain structure and functional dynamics. In our study, we utilize the network control theory to reveal the relationships between the brain's anatomical structure and the observed coactivation pattern of cognitive function.
- Method: We constructed a linear model to model the brain network and then used the pinning control strategy to input energy to the model, forcing the model output to track the real fMRI signals and optimize the node selectionsand input energy. We proposed to use the Half-Quadratic Splitting algorithm to solve the optimized model and analyze the controllability of the brain structure to determine how the brain balances energy consumption and neural circuit integration. We have already tested this framework in working memory task and identify the critical brain control regions and corresponding input energy. Now we are trying to expand the this framework to non-linear model to see if we can get similar results.
- Diagram of the Research:
-
Use GNN to detect and classify the seizure according to the EEG data.
- About: Seizure is a common neuro degenerative disease. The diagnose of seizre is based on the analysis of the EEG data, which is time-consumting and low accuracy. Therefore, in this project, we try to use the deep learning method to improve the process of auto seizure detection and classification from the EEG data.
- Method: To build a network based on the EEG data, it is necessary to include the connection between the EEG channles. Therefore GNN is used in this research to build network model.
-
Use radiomics to predict the pituitary adenoma consistency.
- About: Pituitary macroadenoma consistency can influence the ease of lesion removal during surgery, especially when using a transsphenoidal approach. Unfortunately, it is not assessable on standard qualitative MRI. In this research, we want to use radiomics to solve this problem with MRI images.
- Method: We collect the MRI images of pituitary adenoma of patients and preprocess the images. Then we use Pyradiomics to extract the features from the images and then use the Feature Gradient and the Random Forest to do the feature engineering to produce the optimal feature subset. The feature subset is applied as the input to the fine-tune random forest model to predict the consistency level of the pituitary adenomas in the testing set.
- Diagram of the research:
Here are some of my research papers and articles published in scientific journals and conferences:
-
Reverse engineering the brain input: Network control theory to identify cognitive task-related control nodes
- Authors: Zhichao Liang, Yinuo Zhang, Jushen Wu, Quanying Liu
- Published In: IEEE Engineering in Medicine and Biology Society, 2024
-
Developing a radiomics model to predict tumor consistency of pituitary adenomas using multicenter MRI data
- Authors: Jushen Wu,Pengcheng Wang, Jiayu Xiao, Gabriel Zada, Zhaoyang Fan
- Published: ISMRM 2025
-
Validation of a nonlinear large-scale neural model inversion using magnetic resonance spectroscopy
- Authors: Jushen Wu, Khoi Minh Huynh, Uzay E Emir, Pew-Thian Yap, Guoshi Li
- Published: OHBM 2026
- BEAM: Whole-Brain Excitation-Inhibition Analysis Model based on Functional MRI
🛹Skateboard, 💪Gym, 🎸Guitar, 🎹Piano,🕺Breaking, 🏀Basketball