A Novel Deep Learning Approach for Complete Segmentation of Roots, Soil and Pores in X-ray Tomography Data of Acrylic Resin Embedded Rhizosphere

Correct image segmentation is the pre-requisite for identifying classes of objects in microscopic datasets in order to determine relationships between them. We recently reported on a novel embedding protocol for rhizosphere samples based on the hydrophilic acrylic LR-white resin.¹ X-ray µ-CT data measured on such embedded samples shows only minimal contrast between root and resin which renders segmentation of these data is difficult or even impossible using common methods based on thresholding of histograms or detection of edges.

Here, we demonstrate how this barrier can be overcome using deep learning of convolutional neural networks based on U-Net architecture.² We show successfully segmented roots from resin, where classical machine learning approach Random Forest was not successful in our attempts. Firstly, the embedded samples were characterised by X-ray µ-CT and cut by a water-jet. Roots on the exposed 2D section were identified using epifluorescence and helium ion microscopy. The analysed 2D image plane was then correlated with the X-ray µ-CT data for accurate classification of training 3D image pixels. With a given input image (in this case a greyscale micrograph of resin embedded soil), a trained U-Net model with minimal labelled pixels, semantically segmented the X-ray data set showing roots, soil and pores. Using multiple deep learning algorithms, the U-Net was the most promising architecture to segment rhizosphere X-ray µ-CT and we show the different input parameters which can improve the segmentation process. The deep learning experiment was carried out with the ORS dragonfly image processing software. We show an accurate and fast approach that can be used to segment LR-white embedded rhizosphere X-ray CT data to roots-soil-and pores for further correlative microscopy analysis to interpret complex rhizosphere processes in the future.

Author Contributions: CB embedded the soil samples and trained the deep learning algorithms, Eva Lippold acquired and reconstructed CT data, Matthias Schmidt acquired helium ion microscopy data and discussions on improving data segmentation.

Acknowledgement: This work was conducted within the framework of the priority program 2089, “Rhizosphere spatiotemporal organization-a key to rhizosphere functions” funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project number RI-903/7-1. Funding is acquired by Niculina Musat and Hans Richnow. Authors acknowledge the analytical facilities of the Centre for Chemical Microscopy (ProVIS) at the Helmholtz Centre for Environmental Research, Leipzig, Germany, which is supported by the European Regional Development Funds (EFRE - Europe funds Saxony) and the Helmholtz Association. Authors thank Object Research Systems for providing a free Dragonfly commercial licence for use in this work.

References

• Bandara, C. D.; Schmidt, M.;  Davoudpour, Y.;  Stryhanyuk, H.;  Richnow, H. H.; Musat, N., Microbial Identification, High-Resolution Microscopy and Spectrometry of the Rhizosphere in Its Native Spatial Context. Frontiers in Plant Science 2021, 12 (1195).
• Ronneberger, O., Fischer, P., Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation 2015.