Application of neural networks to the segmentation of nanocrystals and liquid nanodrops during in situ condensation of water in the environmental transmission electron microscope - Archive ouverte HAL Accéder directement au contenu
Poster De Conférence Année : 2023

Application of neural networks to the segmentation of nanocrystals and liquid nanodrops during in situ condensation of water in the environmental transmission electron microscope

S.D. Beye
  • Fonction : Auteur
N. Smigiel
  • Fonction : Auteur
V. Onfray
  • Fonction : Auteur
C. Revol-Muller
  • Fonction : Auteur
M. Duchamp
  • Fonction : Auteur
T. Grenier
  • Fonction : Auteur

Résumé

As most of the scientific research fields, Electron Microscopy benefits from advances in Deep Learning approaches, especially for the segmentation of various phases or nanometric objects to be identified on experimental micrographs. Assisting the human operator helps to quantify large populations of objects evolving under dynamic conditions, i.e. during environmental or in situ experiments [1]. We study here the hygroscopic behavior of model aerosols, i.e. NaCl nanocubes in Environmental Transmission Electron Microscopy (Titan ETEM, FEI/TFS), where water vapor introduced at a few mbar (e.g. 5-15 mbar) can be condensed on the sample once cooled to a few degrees Celsius using a liquid nitrogen cryo-holder (ELSA, Gatan). One interesting question deals with the size effect during deliquescence [2]. Images from video sequences could be processed by considering 3 types of objects possibly present at the same time: ‘dry’ crystals, ‘wet’ crystals surrounded by a more or less extended water layer, and dissolved crystals, i.e. water drops. We have used two types of network architectures. The first architecture refers to 2 UNets [3] to perform separately a semantic segmentation of crystals (‘dry’ or ‘wet’) on the one hand and drops on the other hand. Indeed, the ‘wet’ crystals class are defined as the spatial superimposition of ‘dry’ crystals and drops. The second architecture is a Yolo network [4] able to perform instance segmentation, i.e. detect individual objects in the image even if partially overlapping each other. Crystals inside drops (the ‘wet’ crystals class) can this directly be labeled. The efficiency of neural networks relies strongly on the quality of the training step. In a previous work [1], we developed realistic numeric simulations enabling to define an exact ground truth avoiding tedious and uncertain annotations as performed by human experts. The same approach is used here, see figure 1. UNet was then trained on such simulated images with random crops and data augmentation to prevent overfitting. Yolo was trained similarly but using a pretrained network on a natural image dataset containing 80 classes of objects. We fine-tuned networks adapted to our images with two classes and assessed results on the basis of Dice and F1 scores. References: [1] Faraz, K. et al., Scientific Reports, 12 (2022) 2484. [2] Biskos, G. et al., Aerosol Science and Technology, 40:2 (2006) 97. [3] Ronneberger, O. et al., MICCAI, 9351, 234 (Springer, LNCS, 2015). arXiv: 1505.04597. [3] Wang, C.-Y. et al., CVPR (2023) 7464, arXiv: 2207.02696. Acknowledgements: This work is supported by the French National Research Agency (ANR) through the ‘WATEM’ project 20-CE42-0008. The ETEM work was conducted at CLYM (www.clym.fr). Eric Ehret, Francisco Cadete Santos Aires and Christian George (IRCELYON) are thanked for fruitful discussions. Part of his work was performed within the framework of the LABEX PRIMES (ANR-11-LABX-0063, ANR-11-IDEX-0007 program at Université de Lyon).
Fichier non déposé

Dates et versions

hal-04229761 , version 1 (05-10-2023)

Identifiants

  • HAL Id : hal-04229761 , version 1

Citer

T. Epicier, S.D. Beye, N. Smigiel, V. Onfray, C. Revol-Muller, et al.. Application of neural networks to the segmentation of nanocrystals and liquid nanodrops during in situ condensation of water in the environmental transmission electron microscope. IMC20, Sep 2023, Busan, South Korea. ⟨hal-04229761⟩
3 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More