For my master’s thesis, I created computational models for processing neuroscience data, specifically the massive amount of video data generated by two-photon calcium imaging (2PCI). Neuroscientists use 2PCI to record neuron activity, but manually mapping out individual neurons across terabytes of video makes scalable studies impossible. Existing automated tools either only segment the main body of the cell, or struggle to process lab-grown neural organoids.

To solve this, I built an end-to-end deep learning pipeline, focusing on U-Net segmentation models that I trained to map both the cell bodies and their processes. The biggest hurdle was an extreme lack of training data, as I only had 32 manually annotated images to work with. To get the model to perform, I implemented synthetic image generation (GANs and heuristics), data augmentation, and domain transferability to investigate if a model trained on one type of brain tissue could generalize to another.

I validated the pipeline on two very different datasets provided by the University of Copenhagen. The first involved neural organoids, comparing samples from healthy patients to those with schizophrenia, which allowed me to extract and analyze specific signal features like burst rates and synchronicity. The second dataset contained in vivo recordings of mice to observe how different medications affected their neuronal behavior. The results derived from my models are currently being worked into neuroscience publications.