State-of-the-art in automated image-based species identification systems
5 affiliations
- 1Institute of Medical Microbiology, University of Zurich , ,
- 2Goba Referral Hospital, Madda Walabu University , ,
- 3Department of Clinical Microbiology, Christian Medical College and Hospital Vellore , ,
- 4Department of Radiation Oncology, Inselspital, Bern University Hospital and University of Bern , ,
- 5Department of Health Sciences and Technologies, ETH Zurich , ,
Abstract
Accurate and rapid species-level identification of bacteria and fungi from culture plates are longstanding challenges in diagnostic microbiology. Visual assessment of colony morphology lacks sufficient specificity for definitive identification, while biochemical profiling is often too slow for routine clinical use. Matrix assisted laser desorption ionization - time of flight mass spectrometry (MALDI-TOF MS) has largely replaced biochemical identification in high-income settings because of its speed and accuracy; however, its cost and infrastructure requirements limit accessibility in most low resource settings. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), offers a promising alternative by enabling growth detection and presumptive identification direct from digitalized macroscopic images of culture plates. This review summarizes key aspects of visual identification from agar cultures and recent advances in image-based microbial analysis using CNNs, Siamese networks, object detectors, and Vision Transformers. Automated growth detection and whole-plate interpretation achieved sensitivities of 99.8–100%, specificities of approximately 63–90%, and within-institution AUCs of ≥0.98, although cross-institution performance decreased to ≥0.90; reported workflow benefits included 3–5-hour reductions in time to result and up to 80% less hands-on review of negative cultures. Species-level bacterial classification ranged from a mean sensitivity of 66.2% in an early 18-species model to 90.6% top-1 and 97.6% top-3 accuracy with the contextual DeepColony framework. Fungal models achieved a mean accuracy of 89.7% with FungID and an AUROC of 92.7% with smartphone-based MoldVision, whose F1 score increased from 38.8% during days 1–2 to 92.1% during days 3–5, highlighting the influence of colony maturity, dataset composition, and validation setting. We further discuss the integration of AI image analysis in commercial laboratory automation systems, including Copan WASPLab™, BD Kiestra™ MLA, APAS™ Independence by Clever Culture Systems, as well as the potential of smartphone-based tools in resource-limited settings. Overall, AI-based image analysis can support high-sensitivity screening, workflow triage, colony enumeration, and presumptive identification, but robust external validation, standardized imaging and reporting, regulatory approval, and continued expert oversight remain essential.
Related
- Artificial Intelligence‐Optimized Image Biomarker Identification for Sinonasal Inverted Papilloma Diagnosis Using Image‐Enhanced Versus White‐Light Endoscopy
- Automated quantitative analysis can support the identification of progressive pulmonary fibrosis in fibrotic interstitial lung disease.
- Automated brain MRI image classification through machine learning and deep learning approaches
- Automated acquisition of thyroid cytology image features using self-supervised learning and application to image classification
- Automated correction of training sets for biological image segmentation by convolutional networks
BibTeX
@article{weber2026stateoftheart,
title = {{State-of-the-art in automated image-based species identification systems}},
author = {Lukas Weber and Mohammedaman Mama Hussen and Balaji Veeraraghavan and Eline Meijer and Yukino Gütlin and Alejandro Guerrero-López and Sarah Brüningk and Oliver Nolte and Catherine Jutzeler and Adrian Egli},
journal = {Frontiers in cellular and infection microbiology},
year = {2026},
month = oct,
doi = {10.3389/fcimb.2026.1818023},
url = {https://europepmc.org/article/PMC/PMC13728494},
note = {CC BY},
}
RIS
TY - JOUR
TI - State-of-the-art in automated image-based species identification systems
AU - Lukas Weber
AU - Mohammedaman Mama Hussen
AU - Balaji Veeraraghavan
AU - Eline Meijer
AU - Yukino Gütlin
AU - Alejandro Guerrero-López
AU - Sarah Brüningk
AU - Oliver Nolte
AU - Catherine Jutzeler
AU - Adrian Egli
PY - 2026
DA - 2026/10/20
JO - Frontiers in cellular and infection microbiology
DO - 10.3389/fcimb.2026.1818023
UR - https://europepmc.org/article/PMC/PMC13728494
AB - Accurate and rapid species-level identification of bacteria and fungi from culture plates are longstanding challenges in diagnostic microbiology. Visual assessment of colony morphology lacks sufficient specificity for definitive identification, while biochemical profiling is often too slow for routine clinical use. Matrix assisted laser desorption ionization - time of flight mass spectrometry (MALDI-TOF MS) has largely replaced biochemical identification in high-income settings because of its speed and accuracy; however, its cost and infrastructure requirements limit accessibility in most low resource settings. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), offers a promising alternative by enabling growth detection and presumptive identification direct from digitalized macroscopic images of culture plates. This review summarizes key aspects of visual identification from agar cultures and recent advances in image-based microbial analysis using CNNs, Siamese networks, object detectors, and Vision Transformers. Automated growth detection and whole-plate interpretation achieved sensitivities of 99.8–100%, specificities of approximately 63–90%, and within-institution AUCs of ≥0.98, although cross-institution performance decreased to ≥0.90; reported workflow benefits included 3–5-hour reductions in time to result and up to 80% less hands-on review of negative cultures. Species-level bacterial classification ranged from a mean sensitivity of 66.2% in an early 18-species model to 90.6% top-1 and 97.6% top-3 accuracy with the contextual DeepColony framework. Fungal models achieved a mean accuracy of 89.7% with FungID and an AUROC of 92.7% with smartphone-based MoldVision, whose F1 score increased from 38.8% during days 1–2 to 92.1% during days 3–5, highlighting the influence of colony maturity, dataset composition, and validation setting. We further discuss the integration of AI image analysis in commercial laboratory automation systems, including Copan WASPLab™, BD Kiestra™ MLA, APAS™ Independence by Clever Culture Systems, as well as the potential of smartphone-based tools in resource-limited settings. Overall, AI-based image analysis can support high-sensitivity screening, workflow triage, colony enumeration, and presumptive identification, but robust external validation, standardized imaging and reporting, regulatory approval, and continued expert oversight remain essential.
SN - 2235-2988
ER -