Over the course of this NSF-funded project, we have made substantial advancements in modeling multi-level brain connectivity using cutting-edge artificial intelligence (AI) and machine learning (ML) techniques. Our primary aim has been to develop a novel computational framework that captures brain dynamics through the integration of structural and modality-level neuroimaging data. Through this work, we have achieved notable intellectual merit in developing scalable, interpretable, and clinically relevant computational models for the study of the human brain.

We successfully developed a multi-scale structural connectivity framework that accounts for spatial similarity and temporal correlation across voxel, vasculature, and tissue levels. This has led to improvements in head tissue segmentation using deep neural networks, resulting in enhanced accuracy in the delineation of neuroanatomical structures, especially in aging populations. Our team introduced advanced segmentation models such as DOMINO and DOMINO++ and demonstrated their clinical applicability in neurodegenerative disease assessment.

On the multimodal front, we developed MAGIC, an AI-driven tool that generates contrast-free CT perfusion maps using non-contrast CT images. The tool has been trained on over 13,000 patient scans and evaluated through both quantitative metrics and physician-led subjective analysis. It represents a leap forward in reducing patient risk while maintaining diagnostic accuracy in stroke imaging. These efforts have resulted in multiple publications and ongoing collaborations with external research institutions.

We also made pioneering strides in neuroscience-inspired AI. Notably, we developed and validated models mimicking artificial neuron selectivity and affective conditioning, bridging biological and computational neuroscience. The Visual Cortex-Amygdala model has shown promising results in replicating Pavlovian emotional conditioning, enabling deeper exploration of emotion encoding in AI systems.

Broadening the scope, we developed a large-scale foundation model trained on over 80,000 MRI scans. Leveraging state-of-the-art transformer architectures, the model has outperformed baselines in multiple neuroimaging benchmarks such as BraTS and ADNI, offering a scalable solution for automated analysis of medical images across diverse applications.

Beyond algorithmic advancements, our project had significant broader impacts. We created and curated a multimodal stroke imaging database of over 13,000 de-identified subjects and developed open-source tools to support reproducible research. Our team mentored over 20 students, including undergraduates, master’s, and Ph.D. trainees, many of whom received national recognition, such as the NSF-GRFP, NIH F31, NIH T32, and IEEE NextGen Scholar Awards. Several students were supported through NSF INTERN supplements, promoting industry engagement and translational research experience.

Research outcomes were disseminated through more than 60 publications and presentations at top venues such as MICCAI, RSNA, SfN, BMES, and Brain Stimulation. These efforts have not only pushed the frontier of brain modeling research but also ensured meaningful impact across clinical, educational, and technological domains. As a result, the project has set a strong foundation for next-generation neuroimaging and AI applications in medicine.

Last Modified: 03/28/2025
Modified by: Ruogu Fang