Clinical AI for Stroke Imaging

In this project, I developed a deep learning pipeline to predict final MRI-based stroke lesion masks directly from acute CT perfusion data. The goal was to bridge the diagnostic gap between fast but limited CT imaging and delayed yet more accurate MRI scans in ischemic stroke treatment.

Using the ISLES 2024 multi-center dataset, I implemented a 3D U-Net architecture with a robust MONAI-based preprocessing and training pipeline. Beyond a unimodal baseline, I designed a multimodal early-fusion model combining cerebral blood flow (CBF) and time-to-maximum (Tmax) perfusion maps, significantly improving segmentation stability and boundary accuracy.

The project involved end-to-end ownership, from data normalization and augmentation to model training, evaluation (Dice, Hausdorff), and post-processing, highlighting my interest in applied AI systems at the intersection of healthcare, machine learning, and real-world constraints.