This project focuses on implementing prior spatial knowledge and a random-walk diffusion strategy to simultaneously capture complex structural and functional relationships between brain regions. By applying attention mechanisms for learning multi-view node feature embeddings and dynamically assigning propagation weights, the project enables the identification of significant functional brain network biomarkers and enhances result interpretability. The developed methods will be utilized to analyze real-world brain network data, contributing to a better understanding of the underlying brain mechanisms and assisting in the discovery of essential biomarkers.

Timeline: 2022 - 2023
Location: The Multiscale Bioimaging and Bioinformatics Laboratory (MBB), Tulane University, PI: Dr. Yu-Ping Wang.

In collaboration with the medical and pathology team at Merck UK, this project developed a deep learning-powered AI solution for PD-L1 CPS scoring, inspired by Merck's globally recognized KEYTRUDA® (pembrolizumab) anti-PD-1 therapy. The AI system was fine-tuned using 302 commercial Stained IHC Whole Slide Images, achieving an AUROC of over 0.98, and its clinical significance was evaluated by comparing its performance to pathologist evaluations using data from clinical trials. The AI solution, developed using Python and PyTorch, focused on automated data quality control, cell segmentation, and predictive modeling, with 37,646 annotated cells.

Timeline: 2022 Summer Internship
Location: Biometrics Research and PDL1 Diagnostic Team, Merck & Co., Inc., Supervisors: Dr. Jeong Hwan Kook and Dr. John Kang.

This project aims to integrate multimodal fMRI data to investigate associations among various paradigms and identify key biomarkers. It introduces a multimodal graph-based deep learning approach that incorporates manifold learning, resulting in highly accurate outcomes. The developed methods will be applied to analyze real-world fMRI datasets, providing valuable insights into the relationships between different paradigms and contributing to the discovery of significant biomarkers in neuroscience research.

Timeline: 2020 - 2021
Location: The Multiscale Bioimaging and Bioinformatics Laboratory (MBB), Tulane University, PI: Dr. Yu-Ping Wang.

This project aims to develop advanced graph-based deep learning methods for the analysis of fMRI data and phenotype prediction. By employing semi-supervised graph deep learning with Laplacian regularization, the project addresses the oversmoothing issue, improving model generalization and robustness. The developed methods will be evaluated on real-world fMRI datasets and compared to existing state-of-the-art techniques, with potential applications in clinical practice and neuroscience research.

Timeline: 2019 - 2020
Location: The Multiscale Bioimaging and Bioinformatics Laboratory (MBB), Tulane University, PI: Dr. Yu-Ping Wang.

This project focuses on developing a graph temporal ensembling based semi-supervised convolutional neural network (CNN) to handle noisy labels in histopathology image analysis. A rule-based CNN model, specifically for classifying SCC and ADC breast cancer cells, was designed using TensorFlow and PyTorch frameworks. Advanced preprocessing techniques, such as image augmentation and data normalization, were employed for whole-slide images to minimize potential biases in the dataset. The developed approach aims to improve the accuracy and reliability of histopathology image analysis, contributing to advancements in cancer diagnosis and research.

Timeline: 2017 - 2018
Location: University of Flroida.