Applying machine learning with MobileNetV2 model for rapid screening of vaginal discharge samples in vaginitis diagnosis.

Abstract

Vaginitis is a prevalent gynecological condition that impacts women's quality of life, with most women likely to experience it at least once. Traditional diagnosis involves manually observing vaginal discharge samples under a microscope. This process relies heavily on the technician's expertise and is vulnerable to subjective biases. The study aimed to improve diagnostic accuracy by applying machine learning, specifically the MobileNetV2 model, to automate the classification of vaginal discharge samples. This model supports doctors in identifying causative agents of vaginitis, including Gardnerella vaginalis, fungi, and other pathogens like bacteria or Trichomonas vaginalis. A dataset of 3,164 images from 1,582 vaginal discharge samples of women aged 18 and over was analyzed. Images were taken under a 40x optical microscope with a resolution of 800 × 800 pixels and classified into three groups: Group B (mixed bacteria or Trichomonas vaginalis), Group C (Gardnerella vaginalis, identified by clue cells), and Group F (fungi, e.g., Candida albicans, which appear as hyphae or yeast cells in samples). The model was trained using 80% of data for training, 10% for validation, and 10% for testing. Performance was evaluated using two statistical metrics: the F1 score (a measure of accuracy balancing precision and recall) and the AUC-PR (Area Under the Curve of the Precision-Recall curve, a measure of model reliability for imbalanced datasets). The MobileNetV2 model performed well across all datasets, achieving an F1 score > 0.75 and an AUC-PR > 0.80. It demonstrated the best performance in identifying Gardnerella vaginalis (Group C), with both metrics exceeding 0.90. In conclusion, this study highlights MobileNetV2's potential as a rapid screening tool for vaginitis, particularly in identifying Gardnerella vaginalis (F1 score and AUC-PR > 0.90). While challenges have remained in classifying co-infections (e.g., Groups B vs. F), the model's stability across datasets underscores its practical utility. Integrating AI into vaginitis diagnosis could enhance efficiency, reduce human error, and improve early detection, ultimately advancing patient care.