Scholarlink

Background <p><i>Arabidopsis thaliana</i> is the leading model plant used to study plant-pathogen interactions. High-throughput phenotyping allows for the simultaneous study of many plants with high-frequency image acquisition. Nevertheless, the segmentation of symptomatic plants on natural soil remains challenging, requiring the annotation of hundreds of images and the subsequent training of specialized models for each pathosystem considered. This paper presents a novel approach to segmenting <i>A. thaliana</i> plants’ time series using a single annotated image.</p> Results <p>Images of <i>A. thaliana</i> plants infected with <i>Pseudomonas syringae</i> pathovar <i>tomato</i> strain DC3000 were annotated with precise segmentation masks. We compared various mask segmentation methods; our one-shot learning approach obtained a Dice score of 0.977 on our test dataset. Variables extracted from the segmented images allowed statistical discrimination between infected and control plants. We used our one-shot learning approach without further fine-tuning on a new pathosystem; <i>A. thaliana</i> infected with <i>Ralstonia pseudosolanacearum</i>, strain GMI1000. We obtained a Dice score of 0.966 in the second test dataset. We also obtained a Pearson correlation coefficient of − 0.928 between the annotated quantitative disease index and the variable generated with our method.</p> Conclusions <p>This work provides a pipeline to segment symptomatic <i>A. thaliana</i> plants by leveraging a visual foundation model. The method has been used successfully on two different pathogens, is fast to train, and does not need a large dedicated graphical processing unit. Our method has characterized plant-pathogen interactions of two pathosystems without fine-tuning for the second pathosystem. Its ease of use and low computing requirements should make adapting our approach to other high-throughput phenotyping platforms easy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

OneRosette to predict them all: single plant prompting on a visual foundation model to segment symptomatic Arabidopsis thaliana time series

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月9日

查看全文
  • Felicià Maviane Maciá,
  • Sabine Wiedemann-Merdinoglu,
  • David Rousseau,
  • Nemo Peeters

摘要

Background

Arabidopsis thaliana is the leading model plant used to study plant-pathogen interactions. High-throughput phenotyping allows for the simultaneous study of many plants with high-frequency image acquisition. Nevertheless, the segmentation of symptomatic plants on natural soil remains challenging, requiring the annotation of hundreds of images and the subsequent training of specialized models for each pathosystem considered. This paper presents a novel approach to segmenting A. thaliana plants’ time series using a single annotated image.

Results

Images of A. thaliana plants infected with Pseudomonas syringae pathovar tomato strain DC3000 were annotated with precise segmentation masks. We compared various mask segmentation methods; our one-shot learning approach obtained a Dice score of 0.977 on our test dataset. Variables extracted from the segmented images allowed statistical discrimination between infected and control plants. We used our one-shot learning approach without further fine-tuning on a new pathosystem; A. thaliana infected with Ralstonia pseudosolanacearum, strain GMI1000. We obtained a Dice score of 0.966 in the second test dataset. We also obtained a Pearson correlation coefficient of − 0.928 between the annotated quantitative disease index and the variable generated with our method.

Conclusions

This work provides a pipeline to segment symptomatic A. thaliana plants by leveraging a visual foundation model. The method has been used successfully on two different pathogens, is fast to train, and does not need a large dedicated graphical processing unit. Our method has characterized plant-pathogen interactions of two pathosystems without fine-tuning for the second pathosystem. Its ease of use and low computing requirements should make adapting our approach to other high-throughput phenotyping platforms easy.