Semantic and instance segmentation deep learning methods for nanoparticles detection

Luis A. Jara-Lugo

doi:10.1016/j.mtcomm.2025.112074

Semantic and instance segmentation deep learning methods for nanoparticles detection

2025-04-01

0

PRE

AI

L

Luis A. Jara-Lugo

J

Jesús Caro-Gutiérrez *

F

Félix F. González-Navarro

M

Mario Curiel

J

Jesus A. Armenta-Garcia

O

Oscar Perez‐Landeros

DOI：10.1016/j.mtcomm.2025.112074

原文链接

原文求助

摘要

En 中文

Nanoparticle segmentation is a critical task in nanotechnology, enabling precise characterization of physical properties such as size, shape, and perimeter. This study compares four segmentation methods: U-Net (semantic segmentation), Mask R-CNN (instance segmentation), k-means clustering, and adaptive thresholding, to identify the most accurate and reliable approach. With a dataset of only 102 images, deep learning-based methods were hypothesized to outperform traditional techniques. Performance was evaluated using metrics such as F1 score and segmentation accuracy. U-Net emerged as the top performer across all image categories, achieving an overall accuracy of 96.65% and an F1 score of 0.909. For easy-to-segment images, U-Net achieved an accuracy of 98.95% and an F1 score of 0.951; for medium difficulty, 95.27% accuracy and 0.906 F1 score; and for hard-to-segment images, 92.76% accuracy and an F1 score of 0.87. These results demonstrate the superior accuracy of U-Net over Mask R-CNN, k-means clustering, and adaptive thresholding, highlighting the potential of CNN-based models to significantly enhance nanoparticle characterization and providing a robust foundation for future research.

Keyword:

Nanoparticles

Segmentation

Deep learning

Semantic

Instance

Semantic and instance segmentation deep learning methods for nanoparticles detection

摘要

期刊

机构

引用论文