Neuron Segmentation
A deep-learning pipeline that segments individual neurons in 3D microscopy volumes automatically, turning work that took weeks of manual tracing into hours. Built end to end, from file conversion and cloud storage through GPU inference to in-browser review.
Team
Donggyu Park (lead), Junseong Lee, Eric Shim, Brian Park, Taeyeon Kim, Junho Jang, Juhee Kim — advised by Prof. Min Jee Jang.
Problem
Tracing neurons by hand through a 3D microscopy volume takes weeks, and that labor — not the imaging — is what limits the research. Our datasets are multi-color, with different cells labelled in different hues, so we treated color as a feature the model could learn from rather than information to discard.
What we built
- A 3D U-Net trained on three color channels instead of grayscale, so the model separates touching neurites by color as well as by shape.
- A conversion step that turns raw microscope files into cloud-ready chunks without relying on expensive commercial software.
- Storage on AWS S3 with model inference running on the Illinois Compute Cluster.
- A browser-based viewer where researchers inspect the model output and correct it, feeding those corrections back into training.
Where it stands
File conversion, the cloud upload path and the browser viewer all work, and the model was validated end to end on a public benchmark dataset. Producing enough hand-labelled ground truth is still the bottleneck, and the steps are not yet wrapped into a single web app.
Presented
KSEA Katalyst (February 2026) and MRC 2026.
Acknowledgements
Prof. Min Jee Jang for providing the datasets and advising the project, and Longwei Zhang for guidance on the Illinois Computing Cluster.
Gallery
