Star-detection models count aerosol particles in laboratory images
A peer-reviewed laboratory experiment applied models trained to locate stars to aerosol images without retraining their weights. The models counted salt-water particles emitted by a manikin, with detection totals close to manual counts across 41 images. Human tuning of image settings remained necessary, and the experiment did not establish calibrated particle sizes or airborne pathogen concentrations.
Science··Midday
Astronomical weights transferred to laboratory particles
Models designed to locate stars found aerosol particles in cluttered laboratory images in a peer-reviewed experiment. MobileUNet and ELUNet, two neural networks trained on astronomical imagery, kept their existing weights when applied to aerosols. Researchers illuminated salt-water particles with laser light as a manikin supplied by an artificial lung released them. The collection comprised forty-one images of steady emission and short pressurised plumes, containing 7,123 manually marked particles.[1]
Count totals came close after image tuning
MobileUNet’s detections amounted to 99.15% of the manual count; ELUNet’s amounted to 97.09%. Those ratios compare totals, so missed particles and false detections can offset one another. The researchers also calculated matching-based performance measures. Although the network weights were unchanged, image-normalisation settings were empirically adjusted to bring counts closer to manual annotations. MobileUNet processed the collection faster than ELUNet under the experiment’s computing conditions.[1]
Formal spatial calibration was incomplete, leaving particle dimensions expressed in pixels. The measurements therefore do not establish physical size distributions or airborne particle concentration. Dim particles could disappear into background noise, overlapping objects could merge, and motion streaks could be counted more than once. The experiment tested particle-image analysis with salt water, without human respiratory samples or live pathogens. It did not measure infection risk or protection provided by a particular mask.[1]
