Nugent Score AI 1.0
Model Description
Nugent Score AI 1.0 is an application within the CarbConnect platform. It is a support tool designed to assist microbiologists, researchers, and laboratories in evaluating Nugent scores more efficiently.
By leveraging AI technology, the app simplifies the analysis of Gram-stained images captured through a microscope with a C-mount camera, improving diagnostic consistency and accuracy. The tool is intended to enhance professional microbial analysis workflows and should be used as an aid rather than as a diagnostic device.
This repository contains the model card for the private Hugging Face model repository associated with the application. A companion Space for this app already exists under the CarbGeM organization.
Intended Use and Limitations
Research Use Only (RUO).
This tool is strictly intended for research use only and is not for use in clinical diagnostics or medical procedures. For further details, please refer to the Nugent Score Learn More page on CarbConnect.
Intended users
- Microbiologists
- Researchers
- Laboratories performing microbial image analysis
Out-of-scope use
- Clinical diagnostics
- Medical procedures
- Use as a standalone diagnostic device
Practical limitations
- The model is intended to support expert workflows, not replace professional judgment.
- Evaluation results reported here are based on the described independent test set.
- A confidence score cutoff may classify a subset of images as inconclusive.
How to Use
A video tutorial demonstrating how to use the app is available on YouTube.
Typical workflow:
- Capture a Gram-stained image using a microscope with a C-mount camera.
- Upload the image through the CarbConnect application or its associated workflow.
- Review the AI-assisted Nugent score output and any confidence information provided.
- Use the result as research support only, not as a clinical diagnosis.
Training Data
The advanced bacterial vaginosis (BV) model was developed using a dataset of 1,940 vaginal smear Gram-stained images, including 430 newly collected images.
From this total, 1,230 images were assigned to the training set. To improve performance, RandAugment-based data augmentation was applied across 49 training iterations. This expanded the original 1,230 training images into 60,270 augmented images, which were then used for model training.
Evaluation Results
The performance of the advanced BV model was evaluated using an independent test set of 106 images.
Four-group classification based on Nugent score
- Overall accuracy: 94%
Two-group classification: BV vs. non-BV
- Accuracy: 95%
- Sensitivity: 86%
- Specificity: 100%
Confidence-based filtering
A confidence score cutoff of 0.62 was introduced to improve diagnostic precision.
- 7% of images were filtered out as inconclusive.
- On the remaining 99 images, four-group classification accuracy increased to 96%.
- On the remaining 99 images, two-group classification accuracy increased to 97%.
Citation
If you use or reference this work, please cite:
Naoki Watanabe, Tomohisa Watari, et al. "Performance of deep learning models in predicting the Nugent score to diagnose bacterial vaginosis." ASM Microbiology Spectrum, Vol. 13, No. 1.