Machine Learning in Dynamic Microscopy

Khelina Fedorchuk⁰, Damien Hicks⁰, Sarah Russel⁰, Kajal Zibaei⁰
(0) Swinburne University of Technology

Find me on Tues Nov 24th, 1:40-3pm AEDT in Remo, table 11

Abstract
The tracking of individual, proliferating cells over time has long been more accurately done by eye than by algorithm. Fully automated, simultaneous tracking of multiple cells remains a daunting challenge, particularly for highly motile cells that grow and divide over days or weeks. Here we examine the use of deep convolutional neural networks to track several generations of T-lymphocytes through tens of thousands of time-lapse microscopy images. Neural networksare trained on short sequences of consecutive frames, where time is rendered in the 3rd dimension. This converts the usual 2D detection plus association problem into a single 3D problem. The network is required to identify cells (classification) and determine cell positions (regression). An additional neural network is used to perform high-quality segmentation of each cell in its cluttered environment. These networks work together to provide an automated solution for extracting cell lineage trees from movies of proliferating cells.