Artificial intelligence
Lead data scientist and developer of comprehensive computational workflow
This study presents an automated workflow for analyzing Magnetic Resonance Angiography (MRA) images of cerebrovascular structures in sickle cell anemia (SCA) patients. Previous manual analysis focused only on the largest vessels visible, but small vessels are known to play a role in SCA pathophysiology as well. In this project, quantitative metrics from MRA images were extracted to detect and predict changes in blood vessels as small as 600 µm in diameter that may indicate disease progression. Traditional statistical methods were combined with machine learning techniques including clustering, classification, and Scaled Event-Based Modeling (sEBM), to detect and classify vascular changes in children with SCA and correlate these changes with stroke risk and disease severity.
Some of the key challenges I addressed as a part of this project include:
- Image Segmentation of Complex Vascular Structures: The MRA images contained complex vascular networks that needed to be segmented into small blood vessels while preserving essential details such as vessel shape and pixel intensity.
- High-Dimensional Data Analysis: Processing the large-scale, high-dimensional dataset of vessel metrics (e.g., length, tortuosity, width, etc.) required robust analysis techniques to ensure no important resolution was lost during analysis.
- Class Imbalance: The dataset was imbalanced, as fewer participants experienced severe neurological events, a common issue in biotech and clinical projects. Specialized algorithms were used to rebalance the classes to ensure accurate machine learning classification.
- Clustering and Disease Progression Modeling: Applying machine learning techniques like K-means clustering and sEBM to categorize the patients into meaningful risk groups required optimizing specific algorithms in order to detect disease progression markers and identify the patients at highest risk of adverse events.
- Interpretation of Machine Learning Outputs: Using classification and clustering algorithms to visualize natural groupings within disease stages allowed for visualization of clinical patterns of risk even in relatively small datasets.
Skills developed include: clinical image processing, neuroimaging, machine learning, AI, data balancing, statistical modeling, and metric extraction from complex datasets