Preview

PDF, English - main document Download (27MB) | Terms of use

Abstract

In recent years the parallel progress in high-throughput microscopy and deep learning drastically widened the landscape of possible research avenues in life sciences. In particular, combining high-resolution microscopic images and automated imaging pipelines powered by deep learning dramatically reduced the manual annotation work required for quantitative analysis. In this work, we will present two deep learning frameworks tailored to the needs of life scientists in the context of plant biology. First, we will introduce PlantSeg, a software for 2D and 3D instance segmentation. The PlantSeg pipeline contains several pre-trained models for different microscopy modalities and multiple popular graph-based instance segmentation algorithms. In the second part, we will present CellTypeGraph, a benchmark for quantitatively evaluating graph neural networks. The benchmark is designed to test the ability of machine learning methods to classify the types of cells in an \textit{Arabidopsis thaliana} ovules. CellTypeGraph's prime aim is to give a valuable tool to the geometric learning community, but at the same time it also offers a framework for plant biologists to perform fast and accurate cell type inference on new data.