Interpretable detection-based image classification for differential diagnosis of vaginal discharge using smartphone-assisted microscopy

The vaginal discharge is a common clinical presentation associated with diverse infectious and microbiota-related conditions. Existing artificial intelligence (AI)-based diagnostic approaches have predominantly focused on detection of individual pathogens and have rarely integrated microbiota-associated conditions within a diagnostic framework. We developed an explainable deep learning framework for automated analysis of Gram-stained vaginal smear images acquired using handheld smartphones or tablets positioned over a conventional light microscope eyepiece.

The framework integrates object detection with image-level classification, termed detection-based image classification, to identify diagnostically relevant microscopic structures without requiring exhaustive enumeration of all objects. Five diagnostic categories included vulvovaginal candidiasis (VVC), bacterial vaginosis (BV), trichomoniasis (TV), gonococcal urethritis/cervicitis (GU), and healthy vaginal condition (HL). YOLOv26, RT-DETR and EfficientDet0 were independently trained and evaluated using 1,966 microscopic images comprising 11,069 annotated diagnostic objects. Model explainability and computational benchmarking were assessed. The localization quality was independently evaluated by eight clinicians and microbiologists.

Class-specific analysis demonstrated that TV and GU consistently achieved the highest detection performance across all models (mAP50 = 0.622-0.752), whereas RT-DETR demonstrated superior performance for VVC, BV, and HL. Expert evaluation showed favorable localization, with Good/Excellent ratings in 84.8 and 85.1% of valid evaluations for YOLOv26 and RT-DETR, respectively, with no significant difference between models (OR = 0.97; 95% CI, 0.58-1.61; p = 0.895).

This study presents an accurate and clinically interpretable framework that integrates microbiota-associated conditions and sexually transmitted infections for differential diagnosis of vaginal discharge using conventional light microscopy. By linking image-level predictions to localized diagnostic morphology, the framework provides human-verifiable AI evidence as a proof-of-concept prototype. While promising for low-cost implementation, rigorous prospective multi-center validation is required before broad clinical deployment or autonomous diagnosis.