International Edition
Latest News
Technology

Foundation Models Automate Carbon Nanotube Quantification and Classification in Electron Microscopy

Foundation models cut carbon-nanotube analysis from minutes per image to seconds, a study from the National Institute for Occupational Safety and Health shows. Researchers combined the Segment Anything Model with a DINOv2 vision transformer to segment and classify…

Foundation models cut carbon-nanotube analysis from minutes per image to seconds, a study from the National Institute for Occupational Safety and Health shows.

Researchers combined the Segment Anything Model with a DINOv2 vision transformer to segment and classify four CNT morphologies — fibers, clusters, matrix particles and matrix-surface extrusions — across roughly 1,800 transmission electron micrographs. The pipeline reached 89.5% cross-validated accuracy and 92.7% on a held-out test set while using a fraction of the training data required by earlier convolutional networks, NIOSH reported.

Manual segmentation has long been the bottleneck in occupational exposure assessment. Analysts spend several minutes per image tracing particle boundaries, introducing subjective variability that complicates regulatory comparisons. The new framework reduces that to a single click per region of interest, then masks background noise before the classifier sees the features.

“Accurate characterization of carbon nanotube morphologies in electron microscopy images is vital for exposure assessment and toxicological studies, yet current workflows rely on slow, subjective manual segmentation,” the study authors wrote in the preprint posted to arXiv.

The dual-pathway architecture feeds SAM (ViT-B/16) binary masks into a quantification module for counting and sizing, while DINOv2 (ViT-B/14) extracts multi-scale features from the masked regions only.

Hypercolumn pooling across 12 transformer blocks produces a 3,840-dimensional vector classified by a shallow fully-connected network. Instance-level processing lets the model resolve mixed samples — distinct particle types co-existing in one field of view — that earlier whole-image classifiers could not separate.

NIOSH’s Cincinnati lab supplied the micrographs, which span diverse SEM and TEM imaging conditions. The dataset and code are slated for public release alongside the peer-reviewed publication, the preprint notes.

Why morphology matters for risk assessment

Regulatory agencies distinguish fiber-like CNTs from tangled clusters because aspect ratio and surface chemistry drive different toxicological pathways. The World Health Organization’s fiber paradigm counts particles longer than 5 µm with a 3:1 aspect ratio; bundles and matrix-embedded tubes fall outside that definition but still contribute to lung burden. Automated, morphology-resolved counts let hygienists apply the correct metric to each fraction without re-analyzing the same micrograph.

Congress created the National Institute for Occupational Safety and Health through the Occupational Safety and Health Act, signed by President Richard M. Nixon on December 29, 1970, building on the Division of Industrial Hygiene founded in 1914. The agency operates as part of the Centers for Disease Control and Prevention within the Department of Health and Human Services, distinct from both the National Institutes of Health and OSHA.

Its research laboratories in Cincinnati supplied the micrographs used in this study, part of a national network that includes sites in Morgantown, Pittsburgh, Denver, Anchorage, Spokane, and Atlanta. A professionally diverse staff of 1,200 spans disciplines from industrial hygiene and toxicology to engineering and statistics.

How the pipeline works

  1. Segmentation: SAM generates binary masks from a single positive click per particle. Edge-artifact filtering and noise removal clean the masks before downstream use.
  2. Mask-guided feature extraction: DINOv2 processes the full image, but the SAM mask zeros out background activations so the transformer attends only to CNT pixels.
  3. Hypercolumn pooling: Features from all 12 blocks are concatenated, normalized to zero mean and unit variance, then fed to the classifier.
  4. Classification: A shallow MLP predicts one of four morphology labels per instance.

Performance compared with prior work

The earlier convolutional baseline, reported by the same group in 2022, needed more than 5,000 labeled images and could not separate co-occurring morphologies. The foundation-model approach matches or exceeds its accuracy with roughly one-third the labeled data and adds instance-level discrimination.

Open questions and next steps

The preprint acknowledges that the classifier’s capacity — linear versus shallow MLP — and the choice of hypercolumn layers remain tuning parameters validated only on the current dataset. Generalization to other nanomaterials (graphene platelets, metal-oxide nanowires) and to scanning electron micrographs with heavier charging artifacts has not been tested. NIOSH plans a multi-lab validation study later this year before recommending the pipeline for regulatory submissions.

FAQ

What are the four CNT morphologies the model distinguishes?
Fiber (single tubes), Cluster (tangled bundles), Matrix (tubes embedded in a particle), and MatrixSurface (tubes extruding from a particle surface).
How much faster is the automated workflow than manual segmentation?
Manual tracing takes several minutes per image; the SAM-based tool segments a particle with one click in seconds, and the full pipeline processes a micrograph end-to-end in under a minute on a standard GPU.
Is the code and dataset publicly available?
The authors state the dataset and code will be released upon peer-reviewed publication; as of the preprint date they are not yet public.
Can this method replace manual counting for regulatory compliance?
Not yet. NIOSH intends a multi-lab validation study before endorsing the pipeline for formal exposure assessments.
About the author: Technology Desk

Technology Desk at Archynewsy.