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 FerranteGigaScience 2021 · DOI

TL;DR

This paper introduces ChronoRoot, a novel system combining 3D-printed hardware and deep learning for high-throughput, temporal phenotyping of plant root system architecture. By analyzing root growth over time, ChronoRoot reveals new temporal parameters and insights into plant development, particularly in response to light conditions, which are crucial for understanding plant genetics and adaptation.

What Problem Does This Paper Address?

Understanding plant root development is critical for agriculture and understanding plant responses to environmental changes. However, studying roots is challenging due to the "rhizosphere problem" – the difficulty of accessing and observing roots in their natural soil environment. While existing methods can analyze root system architecture (RSA) at specific time points, temporal phenotyping (observing growth over time) has been hindered by technological limitations. This prevents researchers from capturing dynamic growth patterns and identifying temporal parameters that might be linked to plant genetics or environmental responses, such as circadian rhythms.

Methods

ChronoRoot employs a two-pronged approach:

  • Hardware: A modular, 3D-printed imaging system is described. Each module uses Raspberry Pi, multiple cameras, and infrared (IR) backlighting to capture images of plants growing on agar plates. This setup allows for high temporal resolution imaging (as frequently as every minute) and is designed to be affordable and adaptable.
  • Software (Deep Learning for Segmentation): The core of the phenotyping is a deep learning pipeline for root segmentation.
  • CNN Architectures: The authors explored and proposed several Convolutional Neural Network (CNN) architectures for segmenting roots from images. They compared standard models like UNet, ResUNet, SegNet, and DeepLab with their own proposed models: DSResUNet (Deeply Supervised ResUNet) and an ensemble model combining multiple architectures.
  • Temporal Consistency: A key innovation is incorporating temporal consistency into the segmentation process. This involves using information from previous frames to refine the current segmentation, helping to overcome issues like water droplets or root occlusion.
  • Graph Generation: After segmentation, the root system is skeletonized, and a graph is constructed. This graph represents the root structure, labeling nodes as seed, main root (MR), lateral root (LR), bifurcations, and tips. This graph representation is crucial for extracting detailed phenotypic parameters.

Datasets

The paper describes two datasets:

  • Training/Validation Dataset: Consists of 331 images from 55 videos (approximately 6 images per plate over time), with 11 videos manually annotated by an expert biologist.
  • Testing Dataset: Comprises 55 images from 11 different videos, also annotated by the expert biologist.
  • Use Case Dataset: Used for exploring novel temporal parameters. This dataset includes 12 videos for each of two photoperiod conditions (long day and continuous light), with images captured every 15 minutes. Data from the first 3 days (prior to germination) were discarded.

The authors emphasize that these datasets are available to encourage reproducible research.

Key Insights and Ideas

  • Bridging the Rhizosphere Gap: ChronoRoot offers a practical, low-cost solution to the long-standing challenge of studying root dynamics over time.
  • Temporal Dynamics Reveal Novel Traits: By analyzing growth over time, the system uncovers parameters beyond traditional static RSA measurements, such as the time point when lateral root length surpasses main root length, or oscillations in growth speed.
  • Light-Dependent Root Plasticity: The study demonstrates how different photoperiods (long day vs. continuous light) significantly influence root architecture and growth dynamics, leading to distinct RSA configurations.
  • Clock-Related Root Behavior: The temporal analysis, particularly Fourier decomposition of growth speed, reveals potential links between plant growth rhythms and light cycles, suggesting ChronoRoot can be used to screen for clock-related mutants.
  • Deep Learning for Robust Segmentation: The combination of advanced CNNs and temporal consistency refinement leads to highly accurate and robust root segmentation, even in challenging imaging conditions.

Key Results

  • Segmentation Performance: The proposed DSResUNet model achieved a good balance between speed (less than half a second per image) and accuracy (comparable Dice and Recall to more complex models), with a significantly lower Hausdorff distance. The ensemble model achieved the best performance across all metrics but at the cost of increased processing time.
  • Photoperiod Effects: Plants grown under continuous light (CL) showed faster root growth, a higher density of lateral roots, and a different overall RSA compared to those under long day (LD) conditions.
  • Novel Temporal Parameters: The study identified and quantified new parameters like the "cross-point" (when MR length equals total LR length) and analyzed the dynamics of MR vs. LR growth.
  • Oscillating Growth Patterns: Fourier analysis of growth speed revealed distinct oscillatory patterns related to daily (1/24h) and ultradian (1/12h) frequencies, with differences observed between CL and LD conditions, suggesting a root growth clock.

Limitations and Future Work

  • Hardware Specificity: While the system is designed to be adaptable, changes in lighting conditions or hardware components might require minimal fine-tuning of the deep learning model.
  • Ground Truth Annotation: The paper notes that their dataset only requires binary segmentation masks (root vs. background), which is less demanding than some other methods (like RootNav 2) that require pixel-level annotations for MR and LR separately. However, this also means their current annotation scheme is not directly compatible with models requiring more detailed ground truth.
  • Future Work: The authors suggest that ChronoRoot's ability to reveal novel temporal parameters, especially clock-related features, makes it an ideal platform for identifying genotypes with altered clock traits and for screening clock-related mutants. The modular design also allows for scaling the system based on available funding and experimental needs.

Summary updated

Abstract

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
@inproceedings{gaggion2021chronoroot,
  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}},
  booktitle = {GigaScience},
  year = {2021},
  doi = {10.1093/gigascience/giab052},
}
Figure