Automated cell lineage reconstruction using label-free 4D microscopy

Here we describe embGAN, a deep learning pipeline that addresses the challenge of automated cell detection and tracking in label-free 3D time-lapse imaging. The embGAN requires no manual data annotation for training, learns robust detections that exhibits a high degree of scale invariance and generalizes well to images acquired in multiple labs on multiple instruments.

Methods
Images were acquired using an Olympus IX83 inverted frame equipped with a UPLSAPO60xs2 objective, a Visitech iSIM multipoint confocal scanner, ASI MX2000XYZ stage, and a Hamamatsu Orca Fusion camera. The mCherry channel of JIM113 was acquired using 594 nm excitation and a 605 nm long-pass emission filter using 150 ms exposures and a laser power that was empirically tuned to not cause any qualitative developmental delays versus un-imaged control embryos and maintain a ~100% hatch rate for imaged embryos. Embryos were imaged every 60 seconds with a 750 nm z spacing. DIC images were acquired with the Visitech scanner in brightfield bypass mode, a 50 ms camera exposure and the LED light source tuned to not generate any saturated pixels in the image. DIC illumination was generated using an Olympus UCD8 manual condenser equipped with a U525 oil immersion 1.4 NA top lens and a DICTHR tilt-shift slider. Images were acquired using a micro-manager and cropped and converted to individual tiff volumes using Fiji.