Deep-learning facilitated microscopy for the dissection of durable resistance to plant disease.

Submission Number: 140
Submission ID: 244
Submission UUID: d1529b29-c741-49d0-b54c-45d58e5ea471
Submission URI: /form/project

Created: Tue, 02/22/2022 - 17:26
Completed: Tue, 02/22/2022 - 17:36
Changed: Wed, 06/05/2024 - 14:55

Remote IP address: 67.176.36.130
Submitted by: Anita Schwartz
Language: English

Is draft: No
Webform: Project
Deep-learning facilitated microscopy for the dissection of durable resistance to plant disease.
CAREERS
Figure6_a3.png
deep-learning (303), python (69)
Complete

Project Leader

Jeffrey Caplan
{Empty}
{Empty}

Project Personnel

Randall Wisser, Philip Saponaro, Anita Schwartz
Hening Cui
{Empty}

Project Information

Using existing image datasets from different microscopy platforms, we would like to cross validate and, if needed, retrain the deep-learning model (U-Net CNN) used by DeepXScope for 3D segmentation. Cross validation and retraining of the model will be performed using real data as well as virtual data (fungal infection on host tissue is created by computer simulation). With a validated pipeline, DeepXScope will then be applied to available datasets to generate associated output data for downstream analysis and biological interpretation.

Background: Plant diseases limit the production of crop plants worldwide. Durable forms of natural resistance (genetically determined) are an environmentally friendly and sustainable solution for disease control, but the biology of durable resistance is poorly understood, particularly at the microscale where the pathogens are visible. To better understand pathogenesis, we developed DeepXScope (https://github.com/drmaize/compvision): a deep-learning facilitated pipeline for segmenting 3D microscopy data and quantifying features of host-pathogen interactions. Working in maize (corn) on a fungus causing prior epidemics, DeepXScope is being used to show how macroscopic disease outcomes arise from microscopic events during pathogenesis.

Project Information Subsection

Produce manually annotated ground-truth images to serve as a reference for cross-validation of the existing pipeline. This includes host and pathogen structures segmented by the pipeline. Similarly, produce virtual ground-truth images by running a simulation model of 3D fungal infection in raw image data. Run DeepXScope on existing datasets and estimate accuracy of the pipeline (following methods used previously by our team). Test whether retraining the CNN is required for images from different microscopy platforms. For each dataset, produce graphical and statistical summaries of the quantitative output. Validate and improve the existing README documentation (https://github.com/drmaize/compvision) and create a graphical portrayal of the pipeline.
{Empty}
Graduate student with skills in scientific computing. Should be comfortable command line operations (shell, python) to install and use DeepXScope on a local machine or a high-performance computing cluster. Experience with machine learning is a plus.
{Empty}
{Empty}
{Empty}
University of Delaware
{Empty}
CR-University of Delaware
02/23/2022
Yes
Already behind3Start date is flexible
6 months
07/20/2022
01/18/2023
  • Milestone Title: S1
    Milestone Description: First become familiar with DeepXScope pipeline, including its functions, installation, and how to process images, writing updates to the public README as needed. Focusing on one dataset, generate both manually annotated and simulated ground-truth data. Using these data, use precision and recall metrics to assess accuracy of the pipeline’s segmentation routines. (2 months from start date) Launch presentation.
    Completion Date Goal: 2022-09-20
  • Milestone Title: S2
    Milestone Description: Without alteration, run DeepXScope on existing and incoming datasets to generate standardized output to be shared with the research team. Meanwhile, apply the approach used during S1 to test the pipeline on data collected from two other microscopy platforms (different technology and imaging resolution). Also test the impact of retraining the CNN for the other specific platforms to assess the robustness of the original CNN model. If necessary, use the updated CNN to rerun the corresponding datasets. (4 months from start date)
    Completion Date Goal: 2023-01-02
  • Milestone Title: S3
    Milestone Description: Produce a publication quality graphical schema describing the pipeline and user interaction. Use this to improve the public README documentation. For each dataset, produce graphical and statistical summaries of the quantitative output. (5 months from start date)
    Completion Date Goal: 2023-02-15
  • Milestone Title: S4
    Milestone Description: Ensure all project files generated during the research experience are well organized with internal README documentation. Produce summary documentation with details about methods, results and findings from the research experience. (6 months from start date) Wrap presentation. Exit Interview.
    Completion Date Goal: 2023-01-18
{Empty}
{Empty}
{Empty}
The student will learn how to cross-validate a model for segmenting 3D image data by manual and simulation-based techniques. The student will gain experience in summarising and presenting their results and with constructing a graphical representation for a computational pipeline. The student will be guided by a team using computer vision and microscopy analysis to tackle questions in plant biology.
{Empty}
{Empty}
{Empty}
{Empty}

Final Report

The project led to improvements in a semi-automated computing pipeline (DeepXScope) for image analysis of microscopy data on plant-pathogen interactions. Project developments were grounded in computer science/computer vision.
The use of the DeepXScope pipeline enables extended research in plant science with implications for agriculture.
The project was enabled by existing computational infrastructure and resources at the University of Delaware, Caviness HPC.
In addition to the scientific mentors who provided direction for the project, the research activities engaged multiple personnel including the student who received training and two staff members who helped to develop new tools.
The project was enabled by existing computational infrastructure and resources at the University of Delaware, Caviness HPC.
Not applicable.
Yes. The semi-automated pipeline was adapted to high-performance computing clusters with less system dependencies, making the resource more accessible for other users.
The project improved tools for scientific research that lead to a better understanding of plant-pathogen interactions leading to crop disease. Therefore, the project has indirect impacts on the safeguarding of crop production.
* More manually annotated data should be used to retrain the CNN model underlying the DeepXScope image analysis pipeline
* Additional annotated data should be used for validating results from the pipeline on multiple datasets to evaluate robustness
* Successfully adapted the DeepXScope pipeline to an HPC cluster (Caviness at Univ. Delaware), overcoming several issues with platform dependencies
* The DeepXScope code and README documentation was improved.
* DeepXscope was used to process data from multiple microscopy imaging modalities and parameter settings for segmentation were evaluated.
* Gaps for further development and improvements were determined.