What is it about?
Bacterial vaginosis is a common vaginal condition affecting 23–29% of women worldwide. It is associated with sexually transmitted infections, preterm birth, and neonatal complications. The Nugent score — a grading system based on microscopic examination of Gram-stained vaginal smears — is a standard diagnostic method, but its accuracy depends on the skill and experience of laboratory technicians, leading to variability in results. In this study, we developed convolutional neural network (CNN) models — a type of deep learning artificial intelligence — to predict the Nugent score from vaginal smear images. A total of 1,510 images collected from 151 vaginal discharge specimens at a hospital in Japan between 2021 and 2023 were used for model development. Images were classified into four categories: normal vaginal flora, no vaginal flora, altered vaginal flora, and bacterial vaginosis. Models were evaluated at both low magnification (400x) and high magnification (1,000x). The high-magnification model was further refined into an advanced model and tested against the assessments of two laboratory technicians using an independent set of 106 images. The advanced model achieved 94% accuracy in four-group classification and 95% accuracy in two-group classification (bacterial vaginosis vs. non-bacterial vaginosis), with a specificity of 100%. The average accuracy of the two technicians was 92% in four-group classification.
Featured Image
Photo by Bermix Studio on Unsplash
Why is it important?
Accurate diagnosis of bacterial vaginosis is important for preventing complications, yet access to experienced laboratory technicians is not universal. Our deep learning model matched or exceeded technician-level accuracy, with kappa coefficients of 0.81–0.83 indicating strong agreement between the model and human assessors. Importantly, applying a confidence score cutoff allowed the model to flag only 7% of cases for human review, potentially reducing the microscopy workload to approximately one-fifteenth of the total. This approach could reduce observer variability and support reliable diagnosis in settings where specialist expertise is limited.
Perspectives
Deep learning models have the potential to serve as diagnostic support tools for bacterial vaginosis, complementing rather than replacing human expertise. In this study, the high-magnification (1,000x) model demonstrated superior accuracy and was selected for further refinement into the advanced model. While low-magnification imaging offers practical workflow advantages — such as compatibility with automated microscopy platforms — further validation is needed to confirm its clinical acceptability before it can be recommended as an alternative. Regardless of magnification, further studies using larger and more diverse datasets from multiple institutions are needed before clinical implementation. Our findings contribute to the growing evidence that artificial intelligence can play a meaningful role in clinical microbiology diagnostics.
Dr Naoki Watanabe
Hirosaki University
Read the Original
This page is a summary of: Performance of deep learning models in predicting the nugent score to diagnose bacterial vaginosis, Microbiology Spectrum, November 2024, ASM Journals,
DOI: 10.1128/spectrum.02344-24.
You can read the full text:
Contributors
The following have contributed to this page
