Data Science to Analyze Biomedical Images

Machine Learning algorithms are used very often in the analyses of biomedical images. Deep Learning algorithms are often used to detect the onset, progression and presence of disease using these images. While these algorithms are often successful in terms on prediction accuracy on well-curated datasets, the application of ML models/methods come with its own set of challenges. Firstly, we would like the model to capture biologically relevant insights that increases our understanding of these diseases. Secondly, these models need to generalize in the sense that the performance on one dataset should be observed in other datasets with different data acquisition protocols. Thirdly, these models need to be robust to noise that is inherent to biomedical images. 

In this project, we will work on these issues in the context of functional Magnetic Resonance Images (fMRIs) of the brain. We will use a resting-state fMRI dataset of ADHD and controls. We will predict ADHD / control based on connectivity data. Then, we will evaluate the performance on a related population with ADHD comorbidity. This allows us to assess whether our method identifies patterns related to ADHD symptomology, and how this information transfers between clinical populations. 

Each one of us is unique. Our brains work differently due to both nature and nurture. In other words, differences in genetic makeup, as well our lived experiences makes dynamic brain activity different for each of us. These individual-level signals are called “brain signatures”. Brain signatures are important in the field of precision psychiatry as drug interventions are modeled differently for each patient.

Previously, our research group has used functional MRIs to show that a small group of regions on the brain cortex can be used to explain these individual-specific signals. However, we are now curious to see if we could get similar results with Electroencephalography (EEGs). EEGs have higher temporal resolution, but poorer spatial resolution. Hence, we will have to modify previous approaches to leverage the strengths of EEGs. 

In this project, our goal is to characterize the frequency at which brain signatures can be detected in EEGs. This has applications to 1) devise new AR/VR devices that can incorporate brain signals from EEGs, and 2) guide preprocessing steps in medical and diagnostic uses of EEGs.