Vision Foundation Models Automate Carbon Nanotube Classification in Electron Microscopy

Accurate characterization of carbon nanotube morphologies in electron microscopy images achieves 92.7% accuracy on a held-out test set and 89.5% cross-validated accuracy when utilizing a unified framework that leverages vision foundation models.

This performance outpaces the previous baseline while utilizing a fraction of the training data, solving a labor-intensive bottleneck in exposure assessment and toxicological studies where current workflows rely on slow, subjective manual segmentation that often require several minutes per image.

To overcome the limitations of manual analysis, the dual-pathway architecture integrates the Segment Anything Model and DINOv2. An interactive quantification tool built on the Segment Anything Model (SAM ViT-B/16) segments particles with near-perfect accuracy using minimal user input, generating binary segmentation masks from input scanning electron microscopy images for downstream quantification tasks including counting and sizing.

Meanwhile, a classification pipeline uses the resulting masks to spatially constrain a DINOv2 (ViT-B/14) vision transformer across 12 transformer blocks, extracting features exclusively from particle regions while suppressing background noise and eliminating edge artifacts. By employing hypercolumn pooling, multi-layer features are combined into a 3840-dimensional vector before classification is performed through a fully-connected network.

Evaluated on a dataset of approximately 1800 transmission electron microscopy images spanning diverse scanning electron microscopy and transmission electron microscopy imaging conditions, the framework predicts four morphology classes: Fiber, Cluster, Matrix, or MatrixSurface. Through instance-level analysis, the system successfully untangles mixed samples by accurately identifying different particle varieties present together in the same viewing area.

These findings show that combining self-supervised feature learning with zero-shot segmentation makes nanomaterial evaluation both fast and consistent, turning an inconsistent manual procedure into a reliable, scalable, and data-driven method for evaluating safety risks.

Photo of author

Technology Desk

Technology Desk at Archyde.

5 Netflix Romances Where Love Clashes With Backgrounds and Beliefs