Diagnostic performance of deep learning-based vaginal microecological morphology assessment for bacterial vaginosis...

Routine wet mount microscopy for bacterial vaginosis (BV) and vulvovaginal candidiasis (VVC) is rapid but operator-dependent and lacks ecological context. We evaluated a deep learning-based vaginal microecological morphology assessment (DL-VMM) system integrated in the GE6000 automated vaginal secretion analyzer.

In a prospective cross-sectional study, 500 symptomatic women provided vaginal smears. DL-VMM was compared with routine wet mount microscopy against a composite reference standard (Nugent score, Amsel criteria, fungal microscopy). Key morphological features including clue cells, blastospores, hyphae, trichomonads, leukocytes, epithelial cells, and cleanliness grade were assessed. Diagnostic accuracy and Cohen's kappa were calculated. We also analyzed discordant cases to explore sources of disagreement.

For BV diagnosed by clue cells, DL-VMM achieved 100.00% sensitivity (95% CI: 94.19-100.00%), 98.86% specificity (97.37-99.58%), and κ = 0.95. For VVC by blastospores, sensitivity was 99.32% (95.85-99.96%), specificity 98.87% (96.94-99.70%), κ = 0.97; by hyphae, sensitivity 100.00% (97.68-100.00%), specificity 98.81% (97.04-99.63%), κ = 0.98. Total agreement between DL-VMM and manual microscopy for all parameters ranged from 96.40% to 99.60% (κ 0.89-0.98). No statistically significant differences were found between methods (all P > 0.05).

This prospective diagnostic study demonstrates that the DL-VMM system achieves high diagnostic accuracy for BV and VVC, with substantial to almost perfect agreement with routine wet mount microscopy. The automated system eliminates operator dependence, provides rapid results, and may standardize vaginitis diagnosis, although discordances, though minor, warrant awareness and potential manual review in equivocal cases.