.png){kind=icon}
.png){kind=icon}
Return
Nanoparticle segmentation in electron microscopy images: From systematic dataset preparation to deep generative augmentation
.png){kind=icon}
PRE
AI
V
A
P
A
S
DOI：10.1016/j.mtphys.2026.102111.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Research on nanomaterials depends largely on the accurate characterization of nanoparticles by electron microscopy (EM) images. Low image contrast, complicated particle geometries, and the scarcity of labeled data make automation in this endeavor extremely difficult. While manual labeling takes a lot of time and effort, traditional deep learning models usually require big datasets in order to function reliably. In order to overcome these drawbacks, we propose a three-phase deep learning architecture that can accurately segment data with only few supervised samples. We use a small dataset of 20 micrographs WO 3 nanoparticles acquired by scanning electron microscope (SEM), to create our own dataset, and train the proposed model in the first phase. Despite this small dataset, the model effectively learns to focus on morphologically significant nanoparticle structures while suppressing background noise and imaging artefacts. In the second phase, the trained segmentation model is integrated into a CycleGAN-inspired generative framework to synthesize realistic EM image–mask pairs that reflect the morphological diversity and complexity of actual samples. Finally, these synthetic samples are actively incorporated into the training phase, augmenting the original dataset and substantially improving segmentation performance. By embedding this data generation directly into the learning pipeline, our method achieves robust, morphologically accurate segmentation even in scenarios of extreme data scarcity. Experimental validations confirm notable performance gains, especially for complex particle morphologies with limited ground truth annotations.
Keywords:
nanoparticle segmentation
electron microscopy
deep learning
data scarcity
generative augmentation
Journal
IF:
9.7
Papers:
2.0K
Citations:
1.2W
Organization
No organization information available
Cited Papers