Speaker
Description
Supervised segmentation of nanoparticles in TEM images is held back by annotation: hand-labelling is slow, subjective, and hard to reproduce, the very opposite of FAIR. TEMPOS (Transmission Electron Microscopy Pipeline for Object Segmentation) inverts the problem. Rather than annotate experimental images, it generates physically informed synthetic micrographs whose ground truth is known by construction, trains a Mask R-CNN model purely on that data, and segments real, unseen images, validated on gold nanoparticles and Co₃O₄ nanocrystals The simulation parameters serve as exact, machine-readable metadata, and results are published in FAIR-compliant form using Datasette, alongside a community database for shared datasets.
Beyond the method, this presentation reflects on the process. TEMPOS grew from a side project, begun around 2020–21, into a containerised and openly released tool over roughly five years, a candid case study in open research: what it takes to build reproducible, AI-ready datasets; why "software contribution" proved a truer framing than "novel method"; and how much of the work is social, sharing data across groups, engaging with the open-source community, and learning from the makers of the tools one depends upon.