University of Central Florida
Center for Research in Computer Vision (CRCV)
Research Intern
Dec 2021 – May 2022 · Florida, United States
Cross-registered with my BITS Pilani bachelor's thesis, Towards an end-to-end solution for object segmentation and tracking in biomedical imaging..
Context
Segmenting cells in 3D volumetric microscopy is a problem where models trained on natural images don't transfer. The data is sparse, the boundaries between touching cells are ambiguous, and annotation is expensive enough that datasets stay small. This project set out to establish what current methods could actually achieve on a new dataset of early-stage mouse embryos, and then to do better than that.What I did
- Benchmarked instance segmentation models on a novel 3D dataset of early-stage mouse embryos, establishing baselines the group could build on.
- Designed a U-Net in PyTorch predicting hybrid contour and distance-map representations rather than binary masks directly — giving the network an explicit signal about where cells separate, which is where segmentation usually fails on densely packed volumes.
- Investigated Vision Transformer (ViT) architectures for multi-frame segmentation to capture spatiotemporal context beyond individual time frames.
Publication
Nunley H., Shao B., Denberg D., Grover P., Singh J., Avdeeva M., et al. Nuclear instance segmentation and tracking for preimplantation mouse embryos. Development, 2024, 151(21): dev202817. DOI · Project siteWhat I took from it
- Accuracy held up in early frames and degraded as the embryo grew denser and more occluded. Segmenting each timepoint in isolation throws away exactly the information that resolves those cases — the neighbouring frames. That's what took me toward tracking.
Tech stack
Python
PyTorch
U-Net
Vision Transformers
NumPy
scikit-image
3D volumetric data