← Back to work

Flatiron Institute
Center for Computational Biology · Simons Foundation
Summer Research Associate
May 2022 – Jul 2022  ·  New York, USA
Simons Foundation
Research ML / AI

Context

Reconstructing a cell lineage means following every nucleus through a time-series microscopy sequence and knowing, at each division, which two cells came from which parent. Two things make it hard: annotated training data is scarce, because labelling volumetric sequences by hand is slow, and identity is easiest to lose at exactly the moment it matters most — division. This work approached both, and fed into an ongoing collaboration on tracking in preimplantation mouse embryos.

What I did

  • Addressed the annotation bottleneck by generating synthetic time-series microscopy data with Pix2Pix generative adversarial networks, expanding the usable training set without additional manual labelling.
  • Trained convolutional networks in PyTorch to predict cell cycle stage, probability of division, and average displacement between timesteps — the signals a tracker needs to decide whether a cell has moved, divided, or been lost.
  • Built a two-frame CNN predicting inter-frame displacement fields, mapping nuclei across timesteps to support multi-object tracking and lineage reconstruction over complete sequences.
  • Tackled division specifically: given a parent volume in one frame and candidate daughter pairs in the next, the network learned to accept or reject each association — replacing hand-crafted criteria such as volume, surface area, and inter-cell distance with learned ones.

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 site

A manuscript on the tracking work is in preparation.

What I took from it

Tracking failures are rarely fixed by a better detector. The detector was fine; identity was lost at the association step, where one object becomes two and the usual assumptions about continuity break. Finding the right constraint to impose there mattered more than any architectural change — a lesson that has shaped how I approach tracking since.

Tech stack

Python PyTorch Pix2Pix GAN CNN NumPy scikit-image