{
  "id": 13384823,
  "title": "Label-free identification of bacterial species using adapted vision transformers",
  "url": "https://urgent.news/2026/10/10/label-free-identification-of-bacterial-species-using-adapted-vision",
  "topic": "science",
  "section": "Science",
  "published": "2026-10-10T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.10.09.757828v1?rss=1"
  },
  "original_language": "en",
  "account": "Antibiotic resistance poses a growing global health threat. Accurate identification of bacterial species is crucial for selecting appropriate antibiotics and improving patient outcomes, especially in critical situations like sepsis where time is of the essence. Deep learning models have shown promise in identifying bacteria from phase-contrast time-lapse images of growing cells in microfluidic traps, eliminating the need for fluorescent labels. Previous research demonstrated the effectiveness of this method using laboratory strains. This study builds upon that work by expanding it to clinical bacterial isolates, testing performance on unseen isolates and exploring the potential of autofluorescence as an identifying feature.\n\nTwo validation sets were collected from separate experiments using the same training data. The first set included phase-contrast images of clinical isolate traps harboring five different species: Escherichia coli, Enterococcus faecalis, Klebsiella pneumoniae, Staphylococcus aureus, and Enterobacter cloacae. These images were captured every two minutes. The second set comprised phase-contrast images of clinical isolates of E. coli and K. pneumoniae, isolated from spiked donor blood, captured every minute.\n\nPrevious research employing conventional convolutional neural network (CNN) models yielded unsatisfactory results. Consequently, vision transformers were investigated as an alternative approach. A self-supervised adaptation procedure was devised for pretrained DINOv3 using a frozen teacher, followed by supervised classification using time-lapse montages. The models were trained using either 60 frames (118 minutes) for isolates grown in Mueller-Hinton medium or 55 frames (54 minutes) for blood-isolated isolates. Mean macro-F1 scores were 93.6 {+/-} 1.3% for medium-grown isolates and 87.2 {+/-} 2.7% for blood-grown isolates, representing the average with sample standard deviation across multiple training runs.\n\nTo assess the model's performance over time, the number of frames available to the model during testing was increased gradually. Fourier-domain filtering was employed to simulate a reduced numerical aperture, potentially allowing for a decrease in aperture size without compromising results. While these findings demonstrate the potential of label-free, image-based diagnostics for rapid bacterial identification, further evaluation on independent isolates, experiments, and patient samples is necessary to confirm the generalizability of the method. When validated, this technique could revolutionize the initial treatment of acute bacterial infections by facilitating more informed antibiotic selection.",
  "summary": "Antibiotic resistance is an increasingly serious problem worldwide. Identifying the bacterial species can help guide antibiotic selection and improve patient outcomes. In sepsis, where every minute counts, rapid identification is particularly important, and new approaches are needed. In this study, deep learning models were trained to identify bacterial species from phase-contrast time-lapse…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}