Digging Deeper with AI: Canada-UK-US Partnership for Next-generation Plant Root Anatomy Segmentation

几十年来，植物表型被用来帮助我们理解支撑植物生长和健康的生物学机制。测量植物可以让我们找到产量更高的新作物，或者在面对气候变化或疾病演变时更有弹性的作物。根是植物中看不见的、经常被忽视的部分，但它在作物生长中发挥着关键作用。根系从周围的土壤中吸收水分和养分，一个高效的根系可以改变植物在地上的表现。根系构型（根系在土壤或基质中的排列）的自动测量已经有了大量的研究。来自英国诺丁汉大学、加州萨斯喀彻温大学和美国唐纳德丹福斯植物科学中心的团队开发了获取2D和3D根结构图像的技术，并使用计算机视觉和人工智能软件快速自动测量这些图像。研究根的解剖学——根内细胞的组织结构——已被证明是一项更具挑战性的任务。显微镜和其他类似的根的图像通常是非常高的分辨率，甚至在一个小区域内可能有成千上万的细胞。许多现有的解决方案都专注于2D分割，但与根架构一样，根解剖本身也是一个3D挑战。在我们能够可靠和快速地对这些致密组织结构进行表型分析之前，我们理解根解剖学的生物学机制和益处的能力总是有限的。该项目将利用诺丁汉大学的成像设施，以及诺丁汉、萨斯喀彻温省和唐纳德·丹福斯植物科学中心的世界领先的植物表型和人工智能专业知识，推动支持3D根解剖高分辨率分割的技术。诺丁汉在Hounsfield设施拥有现代化的成像设备：激光消融断层扫描（LAT）和新的微型计算机x射线断层扫描（µCT）平台，以高通量和高分辨率收集3D数据。诺丁汉大学还在牙根解剖的二维分割方面进行了重要的工作，这将为这里开发的三维分割方法提供基础。萨斯喀彻温大学的研究人员是处理大型数据集的专家，他们使用人工智能来检测2D中的物体，以及视频序列中的物体和事件。他们的专业知识将使我们能够在遍历3D堆栈时识别重要的生物特征，并将这些特征与现有的2D分割相结合，形成详细的根组织3D地图。唐纳德·丹福斯植物科学中心的研究人员在植物表型和3D成像以及低成本设备方面拥有专业知识。他们与诺丁汉大学在同一物种上捕获的图像数据将提供重要的跨平台图像可变性，使我们能够训练适用于整个社区的通用模型。通过研究对英国、加拿大和美国经济至关重要的常见作物品种，人工智能解决方案将比单个实验室单独开发的解决方案更通用、更强大。更好地了解根系解剖学将推动生物科学研究，让我们更好地了解根系适应如何影响水和养分的吸收。所有经过训练的模型和最终的分割将与我们在北美的合作伙伴共享，并发布到更广泛的研究社区。

For many decades plant phenotyping has been used to help us understand the biological mechanisms that underpin plant growth and health. Measuring plants lets us seek out new crops that are higher yielding, or more resilient in the face of a changing climate or evolving diseases. Roots, the unseen and often overlooked part of plants, play a pivotal role in the development of strong and robust crops. Root systems extract water and nutrients from the surrounding soil, and a high performing root system can transform the performance of the plant above ground. There has been a great deal of research on the automatic measurement of root architectures - the arrangement of root systems in soil or substrate. Teams including those at the University of Nottingham, UK, the University of Saskatchewan, CA, and the Donald Danforth Plant Science Centre, USA, have developed techniques to acquire images of 2D and 3D root architecture and computer vision and AI software to measure these images quickly and automatically.The study of root anatomy - the organisation of cells within a root - has proven a more challenging task1. Microscopy and other similar images of roots are often very high resolution, and there may be many thousands of cells within even a small area. Many existing solutions have focused on 2D segmentation, but like root architecture, root anatomy is an inherently 3D challenge. Our ability to understand the biological mechanisms and benefits of root anatomy will always be limited until we can reliably and quickly phenotype these dense tissue structures.This project will push forward the technology that underpins high-resolution segmentation of 3D root anatomy by leveraging the imaging facilities at Nottingham, and the world-leading plant phenotyping and AI expertise at Nottingham, Saskatchewan, and the Donald Danforth Plant Science Centre. Nottingham houses modern imaging facilities at the Hounsfield Facility: a Laser Ablation Tomograph (LAT), and new micro-computed X-Ray tomography (µCT) platforms that collect 3D data at high throughput and resolution. Nottingham has also undertaken important work in 2D segmentation of root anatomy, which will provide a foundation for the 3D segmentation methods developed here. Researchers at the University of Saskatchewan are experts in working with large datasets, using AI to detect objects in 2D, and objects and events in video sequences. Their expertise will allow us to identify important biological features as we traverse through the 3D stack, combining these features with the existing 2D segmentations into a detailed 3D map of the root tissue. Researchers at the Donald Danforth Plant Science Center have expertise in plant phenotyping and 3D imaging, and low-cost devices. Their image data captured on the same species as those at Nottingham will provide important cross-platform image variability, letting us train generalisable models that work for the whole community. By working on common crop varieties important to the economies of the UK, Canada and the US, the AI solutions will be more general and more robust than those developed by a single lab working alone.Gaining a better understanding root anatomy will drive forward bioscience research, letting us better understand how root adaptations affect water and nutrient uptake. All trained models and the final segmentations will be shared with our partners in North America and released into the wider research community.