Researchers developed NSYOLO, an AI framework that automatically segments and analyzes nanoparticles in electron microscopy images with high precision, improving accuracy in complex imaging environments and enabling high-throughput nanomaterial characterization.Researchers developed NSYOLO, an AI framework that automatically segments and analyzes nanoparticles in electron microscopy images with high precision, improving accuracy in complex imaging environments and enabling high-throughput nanomaterial characterization.
Nanoparticles are widely used in catalysis, energy storage, optics, biomedicine, and advanced manufacturing because of their unique physicochemical properties. Since nanoparticle performance is strongly influenced by particle size, morphology, and surface structure, accurate characterization of nanoparticles is essential for understanding and optimizing material behavior.
However, analyzing nanoparticles in Scanning Electron Microscopy (SEM) and Transmission Electron Microscopy (TEM) images remains a major challenge. Conventional approaches often rely on manual annotation and measurement, which are labor-intensive, time-consuming, and prone to user-dependent variability. The difficulty becomes even greater when nanoparticles are densely distributed, overlapping, or surrounded by complex background noise.
To address these challenges, researchers developed a new artificial intelligence framework called Nanoparticle Segmentation You Only Look Once (NSYOLO) for automated nanoparticle instance segmentation and morphology analysis.
“Traditional nanoparticle analysis methods often struggle with blurred boundaries, particle overlap, and complicated microscopy backgrounds,” the research team explains. “Our goal was to develop an intelligent system capable of performing accurate and automated nanoparticle characterization with minimal human intervention.”
The proposed framework is built upon the YOLO version 11 (YOLOv11) deep learning architecture and incorporates two advanced modules to improve segmentation performance.
The first module, called Boundary-Aware Dynamic Snake Convolution (BADSConv), enhances the model’s ability to capture irregular and weak particle boundaries. The second module, known as Bi-level Routing Attention (BRA), strengthens global contextual understanding and improves feature extraction in densely populated nanoparticle environments.
The framework was trained using a high-quality SEM/TEM image dataset containing three representative nanoparticle morphologies: nanospheres, nanorods, and nanocubes.
Experimental results demonstrated that NSYOLO substantially outperformed the baseline model and several commonly used open-source nanoparticle analysis tools. The framework achieved a mean Average Precision (mAP@0.5) of 0.957, compared with 0.906 for the original model. The researchers also found that NSYOLO showed particularly strong performance in images containing overlapping particles and complex backgrounds, where traditional methods such as ImageJ and ImageDataExtractor often fail to produce reliable segmentation results.
Beyond the algorithm itself, the team further developed a web-based automated analysis platform that allows users to perform nanoparticle segmentation, particle size statistics, and editable Word report generation without requiring programming knowledge.
According to the researchers, the platform could significantly improve the efficiency and reproducibility of nanoparticle characterization workflows while reducing the workload associated with manual image analysis. Electron microscopy images suitable for processing extend beyond traditional SEM and TEM images to include Bright-Field STEM (BF-STEM) images, Dark-Field STEM (DF-STEM) images, and SEM secondary electron images, provided that the nanoparticles within these images typically display strong intensity contrast and well-defined edges.
“This work demonstrates the potential of combining deep learning with materials characterization,” the authors note. “The proposed framework may help accelerate high-throughput nanomaterial analysis and support future intelligent microscopy systems.”
The researchers believe the framework could also be extended to broader applications involving complex microscopic image analysis, including biological imaging, advanced manufacturing, and automated materials discovery.
Although the current system achieved high segmentation accuracy, the team acknowledges that further improvements are needed for scenarios involving severe particle overlap or extremely weak boundary contrast. Future research will focus on expanding the dataset diversity, improving model robustness, and reducing computational costs through lightweight model optimization techniques.
This paper, “High-precision automated nanoparticle segmentation using a deep learning framework with boundary-aware and attention networks,” was published in AI & Materials .
Han X, Jin L, Zhong Y, Guo C. High-precision automated nanoparticle segmentation using a deep learning framework with boundary-aware and attention networks. AI Mater. 2026(2):0006, https://doi.org/10.55092/aimat20260006.
AI & Materials
Experimental study
Not applicable
High-precision automated nanoparticle segmentation using a deep learning framework with boundary-aware and attention networks
20-May-2026