3D movement tracking for neural activity monitoring

Supervision: Mateusz Koziński, Michał Tyszkiewicz

We are looking to solve a cutting-edge problem in neuroscience: enable recording neuronal activity in a freely-moving C. elegans  worm. We are given sequences of microscopy 3D scans, from which neuronal activity can be measured as the brightness of neuronal nuclei in the scans. The end goal is to enable measurement of activity of selected neurons in a freely moving animal. The main challenge lies in identifying the region corresponding to each of these selected neurons in every frame of the recording. This tracking task is the core of this project.

The video above may only be visible in EPFL VPN.

The tracking is currently performed in a semi-automatic fashion, and is prohibitively expensive in terms of the necessary manual labor. Your job will be to decrease this effort by building an efficient tracker for the 3D microscopy data. We have recently developed two simple, novel, and efficient techniques which use reinforcement learning. One is meant for object tracking, but we have so far only applied it to 2d data (not published yet). The other can be used for finding salient keypoints - it has been tested on problems regarding stereo vision (see the preprint ) but we expect it to be useful for object tracking as well (see this project ). We think that a clever clever combination of these and other techniques can be used to solve the task and greatly increase the volume of data neuroscientists can work with.

Close familiarity with Python and machine learning is necessary, prior exposure to PyTorch will be helpful.