Year: 2026
Type: preprint
Abstract: Low contrast and heterogeneous backgrounds hinder automated molecular detection in time-resolved transmission electron microscopy (TEM). We prepared C60-loaded carbon nanohorns by solvent-free mechanochemistry and compared four preprocessing pipelines using YOLOX detectors trained on expert-annotated frames. Wiener filtering combined with contrastlimited adaptive histogram equalization (CLAHE) yielded the highest F1 score of 0.713 at an intersection-over-union (IoU) threshold of 0.30 on temporally held-out frames from the same sequence. Applied to later frames, the model detected partially overlapping C60 candidates and yielded declining detection counts. Some projected intermolecular separations decreased to 0.90 nm, with a cumulative electron dose on the order of 108 e– nm–2 over approximately 15 s. These observations were consistent with, but not diagnostic of, [2+2] cycloaddition. The model also detected some fused structures as C60, indicating that counts reflect C60-like image features rather than true monomer populations. The workflow provides frame-wise localization and counting of molecular candidates. Multiclass detection of monomers, and fused structures, combined with tracking that accounts for detection errors and independent validation, could support quantitative kinetic analysis observed by atomic-resolution TEM.
Source: ChemRxiv
Language: English
Subfield: Structural Biology
Domain: Life Sciences
Open Access status: green
Awards: G4842022727, G6442224289
OpenAlex ID: W7215037977