The Nugent score is a commonly used diagnostic tool for bacterial vaginosis. However, its accuracy depends on the skills of laboratory technicians. This study aimed to evaluate the performance of deep learning models in predicting the Nugent score to improve diagnostic consistency and accuracy. In total, 1,510 vaginal smear images collected from a hospital in Japan between 2021 and 2023 were assessed. Each image was annotated by laboratory technicians into one of four categories based on the Nugent score-normal vaginal flora, no vaginal flora, altered vaginal flora, or bacterial vaginosis. Deep learning models were developed to predict these categories, and their performance was compared to that of technician annotations. The deep learning models demonstrated 84% accuracy at 400× magnification and 89% at 1,000× magnification. The 1,000× model was further optimized and tested on an independent set of 106 images. After optimization, the advanced model achieved 94% accuracy, outperforming the average 92% accuracy of the technicians. The agreement between the advanced model predictions and technicians was 92% for normal vaginal flora, 100% for no vaginal flora, 91% for altered vaginal flora, and 100% for bacterial vaginosis. Overall, our findings suggest that deep learning models have the potential to diagnose bacterial vaginosis with an accuracy comparable to that of laboratory technicians.IMPORTANCEBacterial vaginosis is a global health issue affecting women, causing symptoms such as abnormal vaginal discharge and discomfort. The Nugent score is a standard method for diagnosing bacterial vaginosis and is based on the manual interpretation of Gram-stained vaginal smears. However, this method relies on the skill and experience of trained professionals, leading to variability in results and poses significant challenges for settings with limited access to experienced technicians. The results of this study indicate that deep learning models can predict the Nugent score with high accuracy, offering the potential to standardize the diagnosis of bacterial vaginosis. By reducing observer variability, these models can facilitate reliable diagnoses, even in settings where experienced personnel are scarce. Although validation is needed on a larger scale, our results suggest that deep learning models may represent a new approach for diagnosing bacterial vaginosis.
Read full abstract