Advancing structural-functional modelling of root growth and root-soil interactions based on automatic reconstruction of root systems from MRI

• Applicants: Professor Dr. Sven Behnke; Professorin Dr. Andrea Schnepf
• Subject Area: Soil Sciences ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2015 to 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 274830790

• Final Report Year: 2021

Final Report Abstract

Root traits are recognized for their importance regarding plant nutrient acquisition; however, their observation in opaque soil is challenging. Recent years have seen strong development in applying 3D imaging techniques such as MRI, µCT or NT to the observation of soil-grown root systems. Starting point of this project was the need of a robust automated algorithm that could provide both the segmentation and reconstruction of those root systems in an automated way. We were particularly interested in using the reconstructed root system architectures to perform simulations of water flow and nutrient transport on these geometries. For that, not only the root length but also the root system topology is of great importance. This has been a limitation of existing automated reconstruction algorithms. Manual reconstruction on the other hand is very time consuming and, thus, a major bottleneck in the overall workflow. In this project, we developed a new two-step method for the automated segmentation and reconstruction of root architectures from MRI images. To improve both contrast and resolution of MRI, we adapted the state-of-the-art deep learning semantic segmentation methods RefineNet and U-Net for 3D super-resolution segmentation of plant root MRI images to segment roots from soil. The second stage of the fully automated algorithm finds the largest component connected to the shoot generated in the first stage and uses 3D skeletonization to reconstruct the root graph structure. The structural reconstruction was implemented using the Python programming language and Mayavi for visualization. The root vs. soil segmentation was trained on synthetic MRI data generated from manual reconstructions and generated root structures. It was applied to reconstruct real MRI images from earlier work and measured in this project. We performed experiments with lupine plants grown in two different substrates using the 4.7 T super wide bore MRI scanner available at IBG-3, Forschungszentrum Jülich. Images were acquired using multislice multi echo imaging sequence (MSME) with a single echo read-out. Selected samples were sent to the µVIS Center at the University of Southampton for comparative µCT scans. The structural reconstruction also proved successful on segmented µCT images of faba bean and maize root systems provided to us by UFZ Halle. Altogether, or algorithm has been evaluated on 24 root systems from two different image acquisition methods and 3 different plants. In general, the number of roots found increases with each automated step. The learned segmentation finds thin roots missed by human expert annotators, which already improves the manual structure reconstruction. In some cases, the automated reconstruction does not result in the correct topology, compared to human annotation. In these cases, the result of the automated reconstruction can serve as input to a manual reconstruction workflow, e.g. in a virtual reality system using head-mounted displays. Our goal for the future is to develop an iterative workflow in which automated reconstruction results can be corrected by humans in a suitable 3D virtual environment in such a way that this further improves the training of the automated algorithm.

Publications

• (2019). Learning super-resolution 3D segmentation of plant root MRI images from few examples. In 27th European Symposium on Artificial Neural Networks (ESANN)
  Uzman, A. O., Horn, J., and Behnke, S.
  
• (2019). Mechanical and hydric stress effects on maize root system development at different soil compaction levels. Frontiers in Plant Science, 10:1358
  de Moraes, M. T., Debiasi, H., Franchini, J. C., Bonetti, J. d. A., Levien, R., Schnepf, A., and Leitner, D.
  
• (2020). 3D U-Net for segmentation of plant root MRI images in super-resolution. In 28th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN)
  Zhao, Y., Wandel, N., Landl, M., Schnepf, A., and Behnke, S.
  
• (2020). Call for participation: Collaborative benchmarking of functional-structural root architecture models. The case of root water uptake. Frontiers in Plant Science, 11:316
  Schnepf, A., Black, C. K., Couvreur, V., Delory, B. M., Doussan, C., Koch, A., Koch, T., Javaux, M., Landl, M., Leitner, D., Lobet, G., Mai, T. H., Meunier, F., Petrich, L., Postma, J. A., Priesack, E., Schmidt, V., Vanderborght, J., Vereecken, H., and Weber, M.
  
• (2020). Functional-structural modelling of root water uptake based on measured MRI images of root systems. In EGU General Assembly 2020
  Selzner, T., Landl, M., Pohlmeier, A., Leitner, D., Vanderborght, J., and Schnepf, A.
  
• (2020). Functional-structural modelling of root water uptake based on measured MRI images of root systems. In Kahlen, K., Chen, T.-W., Fricke, A., and Stützel, H., editors, FSPM2020: Towards Computable Plants, page 127
  Selzner, T., Landl, M., Pohlmeier, A., Vanderborght, J., Leitner, D., and Schnepf, A.
  
• (2020). Reconstructing root system architectures from non-invasive imaging techniques for the use in functional structural root models. In EGU General Assembly 2020
  Landl, M., Huber, K., Pohlmeier, A., Vanderborght, J., Pflugfelder, D., Roose, T., and Schnepf, A.
  
• (2021). Robust skeletonization for plant root structure reconstruction from MRI. In 25th International Conference on Pattern Recognition (ICPR)
  Horn, J., Zhao, Y., Wandel, N., Landl, M., Schnepf, A., and Behnke, S.