Introduction

Medical imaging plays a central role in modern healthcare; however, many advanced imaging techniques remain expensive, time-consuming, and difficult to deploy in routine clinical settings. Magnetic resonance imaging (MRI) is a non-invasive technique for assessing brain health. Functional MRI (fMRI) measures brain activity by detecting blood flow changes, revealing which brain regions are active during tasks or at rest. While fMRI provides rich information about brain function, it presents significant acquisition challenges: scans are lengthy, require active participant engagement during tasks, and are highly susceptible to motion artifacts and inter-individual variability, particularly problematic for individuals with neurological or movement disorders. In contrast, structural MRI (sMRI) captures the brain's physical anatomy, its shape, size, and tissue structure. SMRI scans are faster, produce more consistent signals, are widely available, and are routinely collected in hospitals worldwide. This discrepancy presents an important opportunity for machine learning (ML): can functional brain activity be inferred from structural brain images alone? Addressing this question could significantly reduce reliance on complex functional scans while still enabling access to meaningful functional information.

Project Description

This project builds on an ongoing neuroimaging and ML study aimed at predicting task-evoked fMRI activation patterns from structural MRI scans alone.

The project uses brain eigenmodes, mathematical representations that capture the natural spatial patterns of the brain's cortical surface [1]. Similar to how a musical instrument vibrates in specific harmonic modes, or how Fourier transforms decompose images into frequency components, eigenmodes decompose the brain's 3D geometry into fundamental spatial patterns. Each eigenmode represents a distinct geometric pattern: some capture large-scale, smooth variations across the entire brain, while others represent fine-grained, localized features. Any brain scan can be expressed as a weighted sum of these eigenmodes. The weights from structural MRI (derived from cortical thickness map) serve as our structural features, while weights from task-fMRI (derived from brain activation maps during specific tasks) serve as our functional features. The core ML  challenge is to train models that predict functional weights from structural weights, effectively inferring brain functional features from brain structure.

In the first phase, students will 

• Generate baseline results with the current prediction pipelines and with the Human Connectome Project (HCP) dataset [2], which provides high-quality structural MRI and multiple task-evoked fMRI scans from approximately 1,200 healthy participants 
• Benchmark against other machine learning and deep learning (DL) models  for structure-to-function prediction

The second phase extends the baseline work by incorporating additional structural feature representations, refining feature selection strategies and conducting performance evaluations with emphasis on generalizability, interpretability and robust assessment. 

Figure 1. Structure-to-function prediction workflow.

Impact

This project lies at the intersection of machine learning and healthcare. By enabling the prediction of functional brain information from widely available structural MRI scans, the work has the potential to improve accessibility to functional insights, reduce patient burden, and support future clinical decision-making. While the project is grounded in healthy cohort data, the developed models and methodologies are aimed to be readily transferable to neurological and neurodegenerative disease applications.

Skills Developed

Students will gain hands-on experience in applied ML/DL using real-world medical imaging data. Key skills developed include conducting research on a challenging and impactful healthcare problem, Python-based ML/DL development, feature engineering, model evaluation and validation, and reproducible research practices. The project also offers exposure to interdisciplinary research at the intersection of AI and healthcare, equipping students with valuable skills for careers in AI, data science, and health technology.