ChronoRoot: High-throughput phenotyping by deep segmentation networks reveals novel temporal parameters of plant root system architecture

Nicolás Gaggion, Federico Ariel, Vladimir Daric, Éric Lambert, Simon Legendre, Thomas Roulé, Alejandra Camoirano, Diego H. Milone, Martín Crespi, Thomas Blein, Enzo FerrantebioRxiv 2020 · DOI

TL;DR

This paper introduces ChronoRoot, a low-cost, 3D-printable system that combines deep learning for image segmentation with high-temporal-resolution imaging to analyze plant root growth. It enables the extraction of novel temporal parameters, offering new insights into plant development and genetics.

What Problem Does This Paper Address?

Understanding plant root development is crucial for improving crop yields and understanding plant adaptation to environmental changes. However, studying roots is challenging due to their location beneath the soil surface (the rhizosphere). Traditional methods for analyzing root system architecture (RSA) are often manual, time-consuming, and limited to single time points, missing the dynamic growth processes. While automated phenotyping platforms exist, they can be expensive and inaccessible. This paper addresses the need for an affordable, high-throughput, and temporally resolved method for plant root phenotyping.

Methods

ChronoRoot employs a multi-stage approach:

  • 3D-Printed Hardware: A modular, open-hardware system is designed and 3D-printed, allowing for the imaging of up to four plants simultaneously. This system uses off-the-shelf electronics (Raspberry Pi, cameras) and infrared backlighting for consistent imaging day and night.
  • Deep Learning for Segmentation: A novel deep learning model, specifically a Deeply Supervised ResUNet (DSResUNet), is developed for segmenting plant roots from image sequences. This model leverages residual blocks and deep supervision for improved accuracy and efficiency. An ensemble of multiple CNN architectures is also proposed for even higher accuracy.
  • Temporal Consistency Refinement: To improve segmentation robustness, the system incorporates temporal consistency by averaging segmentations across consecutive frames, helping to mitigate issues like water droplets or temporary occlusions.
  • Graph-Based RSA Reconstruction: The segmented root images are converted into labeled graphs. This graph representation captures the main root (MR), lateral roots (LRs), and bifurcations, allowing for detailed analysis.
  • Feature Extraction: Both traditional RSA parameters (e.g., MR length, LR number) and novel temporal parameters are extracted from the reconstructed root graphs. These novel parameters include metrics derived from spectral analysis of growth speed signals.

Datasets

The study utilized two datasets:

  • Training and Validation Dataset: Consisted of 331 images from 55 videos, with 11 videos manually annotated by an expert biologist.
  • Use Case Dataset: Used for plant phenotyping under alternative photoperiods, comprising 12 videos for each condition, with images captured every 15 minutes. The first 3 days (prior to germination) were discarded, and 25 plants per photoperiod were analyzed.

The paper emphasizes that these datasets are made available to encourage reproducible research.

Key Insights and Ideas

  • Affordable High-Throughput Root Phenotyping: The combination of 3D printing and deep learning offers a cost-effective solution for detailed root analysis, democratizing access to advanced phenotyping capabilities.
  • Temporal Dynamics Reveal Hidden Traits: Analyzing root growth over time, rather than at single points, uncovers dynamic behaviors and allows for the identification of novel parameters that are missed by traditional methods.
  • Deep Learning for Robust Segmentation: The proposed DSResUNet and ensemble methods demonstrate significant improvements in root segmentation accuracy, even in challenging conditions like occlusions.
  • Spectral Analysis of Growth Speed: Applying Fourier analysis to root growth speed reveals underlying oscillatory patterns, potentially linked to circadian or ultradian rhythms, offering a new way to study plant clocks.
  • Light Environment Impacts Root Architecture: Different light conditions (continuous vs. long day) lead to distinct root system architectures, highlighting the plasticity of root development in response to environmental cues.

Key Results

  • Segmentation Performance: The proposed DSResUNet achieved a good balance between speed (under 0.5 seconds per image) and accuracy, with a low Hausdorff distance. The ensemble method further improved accuracy at the cost of increased processing time.
  • Temporal Parameter Discovery: ChronoRoot enabled the identification of novel temporal parameters, such as the time point when lateral root length surpasses main root length, and spectral features of root growth speed.
  • Photoperiodic Effects: Plants grown under continuous light showed faster overall root growth and a different RSA, characterized by a higher density of lateral roots and a lower proportion of the main root compared to plants under long-day conditions.
  • Clock-Related Disorder Detection: Spectral analysis of root growth speed revealed differences in oscillatory behavior between light conditions, suggesting ChronoRoot's potential for screening clock-related mutants.

Limitations and Future Work

  • Manual Annotation Requirement: While the system automates acquisition and analysis, the initial training of the deep learning models still requires manual annotation of root images, which can be labor-intensive.
  • Generalizability to Different Species/Conditions: The current models are trained on Arabidopsis thaliana. Further retraining or fine-tuning might be necessary for other plant species or different growth media (e.g., soil).
  • Future Work: The authors suggest that ChronoRoot can be applied to studies of natural variation, genetics, and screening for clock-related mutants. Further development could involve expanding the analysis to other plant species and exploring more complex environmental interactions.

Summary updated

Abstract

Abstract Background Deep learning methods have outperformed previous techniques in most computer vision tasks, including image-based plant phenotyping. However, massive data collection of root traits and the development of associated artificial intelligence approaches have been hampered by the inaccessibility of the rhizosphere. Here we present ChronoRoot, a system which combines 3D printed open-hardware with deep segmentation networks for high temporal resolution phenotyping of plant roots in agarized medium. Results We developed a novel deep learning based root extraction method which leverages the latest advances in convolutional neural networks for image segmentation, and incorporates temporal consistency into the root system architecture reconstruction process. Automatic extraction of phenotypic parameters from sequences of images allowed a comprehensive characterization of the root system growth dynamics. Furthermore, novel time-associated parameters emerged from the analysis of spectral features derived from temporal signals. Conclusions Altogether, our work shows that the combination of machine intelligence methods and a 3D-printed device expands the possibilities of root high-throughput phenotyping for genetics and natural variation studies as well as the screening of clock-related mutants, revealing novel root traits.
BibTeX
@misc{gaggion2020chronoroot,
  author = {Nicolás Gaggion and Federico Ariel and Vladimir Daric and Éric Lambert and Simon Legendre and Thomas Roulé and Alejandra Camoirano and Diego H. Milone and Martín Crespi and Thomas Blein and Enzo Ferrante},
  title = {{ChronoRoot: High-throughput phenotyping by deep segmentation networks reveals novel temporal parameters of plant root system architecture}},
  year = {2020},
  doi = {10.1101/2020.10.27.350553},
  howpublished = {bioRxiv},
}
Figure