From Acquisition to Answers: Developing Open-Source Automation and Quantitative Analysis Workflows with Built-In Statistics for Electron Microscopy

Le résumé fourni par la source

Automated electron microscopes are making it routine to collect large volumes of data at a pace that outstrips manual interpretation. This gap is especially evident when experiments span heterogeneous samples, variable contrast, noisy data, and subtle or sparse nanoscale features. Here we present an open-source, AI-enabled analysis stack that converts high-throughput acquisition into quantitative measurements with built-in statistical reporting across diverse electron microscopy datasets. The stack emphasizes ease of use, transparent model choice, and reproducibility for both expert and novice users. We first highlight our recent low-dose cryoTEM work using YOLO-based segmentation to extract quantitative ultrastructural measurements from challenging micrographs of vitrified Gram-negative bacteria. Although demonstrated on micron-scale cells, the broader takeaway is the ability to detect and measure fine, sparse, low-contrast features at the nanometer scale (e.g., nm-diameter flagella and nm-level membrane thickness) and to maintain performance across heterogeneous fields of view in non-purified samples by explicitly separating true signal from common confounders such as grid carbon edges and contamination. Coupled with TB-scale data collection, these workflows enable statistically powered measurements across heterogeneous populations rather than small, manually curated subsets. They also include automated distribution summaries and outlier-aware feature statistics to help end users interpret complex samples. Next, we introduce a web-first analysis environment built around a two-prong segmentation strategy for higher–signal-to-noise EM modalities. The first prong is a SAM3-based workflow with curated presets for recurring targets in SEM (e.g., nanoparticles, spores, vesicles, and other defined feature classes), enabling rapid, repeatable mask generation with minimal tuning and minimal user input. The second prong is a trainable, clickable model designed for fast adaptation: users correct a small number of predictions interactively, and the model learns dataset-specific cues to improve segmentation consistency across large batches. With appropriate modifications, these workflows can also be applied to select STEM datasets. The backend is designed to run on single- or multi-GPU systems, while users interact through an accessible web-browser front end that supports batch processing and standardized exports. Finally, we present a model-selection, statistics, and reporting workflow that helps users identify the most reliable model for quantitative analysis without requiring expertise in model training or evaluation. Users can choose among candidate models (e.g., U-Net, YOLO, SAM-derived approaches) [1-4] based on task-relevant metrics, detection confidence, computational cost, and training time. The system automatically produces training/validation summaries, error audits, and measurement-level statistical outputs (e.g., feature distributions, confidence intervals/bootstraps where appropriate, between-condition comparisons, and QC flags) that make results traceable across the experiment. We demonstrate this benchmarking approach under low-dose (<100 e/Ų) and ultra-low-dose (<10 e/Ų) cryoTEM conditions, where SNR and artifact sensitivity can strongly shift model rankings and where explicit, metric-driven model choice is critical before reporting measurements. Together, these components provide an open-sourced practical path from automated collection to automated, trustworthy quantification, enabling users to extract measurements and statistics from heterogeneous EM datasets efficiently, with minimal microscopy- or workflow-development expertise [5].

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.