Answers are grounded in this paper and cite their sources.
Quantifying vesicles and nanoparticles in cryo-electron microscopy (cryo-EM) micrographs is often limited by low signal-to-noise ratios [1], defocus artifacts from intentional underfocus [2], heterogeneous ice thickness [3], and residual beam-induced motion that fragments or distort particle rims [4,5]. In crowded fields, vesicles frequently touch or overlap in 2D projection, so intensity-based masks and morphology operations merge neighbors, while ice texture, filament-like debris, and aggregates create false positives [6]. These errors bias size and shape distributions used for microstructure–property interpretation and formulation decisions [7]. We introduce NanoRange, an end-to-end, instance-aware workflow Figure (1), that converts raw cryo-EM micrographs into per-object measurements by combining Gemini-based AI agents with deterministic image processing and geometric fitting, Figure (2a, 2b). The design goal is to keep the pipeline both adaptive (so it can handle changing SNR, background, and ice conditions across micrographs) and transparent (so each step produces interpretable intermediate outputs that can be audited and tuned). Rather than relying on a single monolithic model, NanoRange uses multiple specialized agents to address distinct failure modes, parameter sensitivity, weak rim contrast, and merge/split errors in crowded regions, while classical processing steps enforce stability and reproducibility. An AI parameter optimizer selects a compact set of enhancement and segmentation hyperparameters per micrograph. An AI contrast enhancer improves rim continuity, followed by adaptive thresholding and conservative cleanup to suppress speckle while limiting unintended mergers. A key step is AI boundary colorization, which produces instance-separated, color-coded contours in regions where multiple vesicles touch or overlap. Each instance is then fit with least-squares circle/ellipse models to extract equivalent diameter, major/minor axes, aspect ratio, and fit-residual quality metrics. The pipeline exports annotated overlays and structured outputs (PNG/CSV/HTML) and support optional human-in-the-loop refinement. NanoRange is evaluated on two liquid fabric enhancer cryo-EM datasets that stress traditional pipelines: one dominated by nonuniform background and filament-like features, and one dominated by dense overlap across a broad size distribution, as seen in Figure (2a). In both cases, instance separation stabilizes geometric fitting and yields more reliable size/shape statistics for downstream distribution analysis and rapid semi-automated labeling. NanORange analysis workflow NanoRange example. (b) Raw cryo-EM micrograph. (c) Instance-separated, color-coded boundaries used to prevent merge/split errors prior to circle/ellipse fitting.
No takes yet. Share an insight, caveat, or question.
Golsorkhi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: