Can U-Net replace CLV3? Machine learning for the identification of biological landmarks in Shoot Apical Meristem images - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Poster Année : 2022

Can U-Net replace CLV3? Machine learning for the identification of biological landmarks in Shoot Apical Meristem images

Résumé

Main engine of plant phyllotaxis, the Shoot Apical Meristem (SAM) is a tightly regulated tissue that presents striking spatiotemporal periodicity properties, and due to that, a high level of inter-individual similarity at tissue scale. It is possible to take advantage of this shape similarity to align a population of SAMs imaged using confocal microscopy onto a common reference frame. In previous work, we performed such an alignment by identifying key biological landmarks on the surface of the SAM. Among these landmarks, an essential one is the precise position of the center of the Central Zone (CZ) of the SAM. It is usually accessed by adding a transcriptional reporter for the CLAVATA3 peptide (CLV3) in the genetic construct of the observed plants. However, the crossing and selection of plant lines generally involves a significant amount of time. In this work, we study whether the geometry of the SAM itself could be sufficient to predict accurately the location of the CZ, and therefore avoid the time-consuming development of crossed reporter lines. We used 3D confocal SAM images containing both a CLV3 fluorescent reporter and a geometry reference marker (either a nuclei-targeted marker under a ubiquitous promoter, or an external cell wall staining). We trained various Machine Learning approaches on both image and surface mesh data to predict the CZ membership based only on geometrical cues. The best performing method among those we tested was the convolutional neural network model U-Net applied on downsampled 2D projections of the reference images. The prediction classifies image pixels into 3 classes: background, meristem or central zone from which it is possible to derive a 2D position for the center of the CZ. We evaluated the different methods on an independent set of images of wild-type SAMs grown under the same conditions, and we show that the trained 2D U-Net model is able to position the CZ center with an error of less than 1 cell relatively to the CLV3 center. It outperforms the surface-based methods, including the recently proposed paraboloid approximation of the SAM surface. Moreover, although the performance drops on images of tilted meristems, it proves to be robust to different growth conditions and to the type of geometry marker used. This tool could constitute a first step towards the aggregation of quantitative SAM data from various experiments without having to rely on CLV3.
Fichier principal
Vignette du fichier
SAM_Poster.pdf (2.72 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03892383 , version 1 (09-12-2022)

Identifiants

  • HAL Id : hal-03892383 , version 1

Citer

Anthony Scriven, Carlos Galvan-Ampudia, Guillaume Cerutti. Can U-Net replace CLV3? Machine learning for the identification of biological landmarks in Shoot Apical Meristem images. From genes to plant architecture: the shoot apical meristem in all its states, Nov 2022, Poitiers, France. ⟨hal-03892383⟩
46 Consultations
14 Téléchargements

Partager

Gmail Facebook X LinkedIn More