Towards Automatic Embryo Staging in 3D+T Microscopy Images using Convolutional Neural Networks and PointNets

TL;DR

This paper compares CNN and PointNet-based methods for automatic embryo staging in 3D+t microscopy images, demonstrating their effectiveness with deviations of 21-34 minutes and a proof-of-concept showing less than 7 minutes deviation.

Contribution

It introduces and evaluates two novel approaches for automatic embryo staging using deep learning on 3D microscopy data, including a point cloud-based method.

Findings

Both methods achieve accurate staging with 21-34 minutes deviation.

PointNet approach shows potential with less than 7 minutes deviation in simulated data.

Methods are suitable for automatic embryo development analysis.

Abstract

Automatic analyses and comparisons of different stages of embryonic development largely depend on a highly accurate spatiotemporal alignment of the investigated data sets. In this contribution, we assess multiple approaches for automatic staging of developing embryos that were imaged with time-resolved 3D light-sheet microscopy. The methods comprise image-based convolutional neural networks as well as an approach based on the PointNet architecture that directly operates on 3D point clouds of detected cell nuclei centroids. The experiments with four wild-type zebrafish embryos render both approaches suitable for automatic staging with average deviations of 21 - 34 minutes. Moreover, a proof-of-concept evaluation based on simulated 3D+t point cloud data sets shows that average deviations of less than 7 minutes are possible.