AI-assisted diagnosis of vulvovaginal candidiasis using cascaded neural networks.

Abstract

Vulvovaginal candidiasis (VVC) is a prevalent fungal ailment affecting women globally. Timely and accurate diagnosis is crucial. Traditional methods, relying on clinical evaluation and manual microscopic examination, have limitations. Artificial intelligence (AI) offers potential improvements in diagnostic accuracy and efficiency by objectively analyzing microscopic images of lower genital tract infections in women. A cascaded model was developed using 100,387 microscope images and 1,761 slides to diagnose VVC at slide level. The model's diagnostic accuracy was compared with experts'. Five hundred thirteen slides were used to evaluate whether the experts' diagnostic skills could be improved using the model as an AI-assisted tool. The consistency between experts' interpretations of microscopic digital images and microscopic examination under eyepiece was assessed to determine whether the collected images adequately represented the slides. The model obtained AUC = 0.9447, 0.9711, and 0.9793 for slide-level diagnosing yeast hyphae, budding yeast, and yeast. Compared with the average performance of experts, the Youden indexes of our model's best points were improved by 0.0069, 0.0772, 0.0579, and 0.0907 for yeast hyphae, budding yeast, yeast, and VVC. The average accuracy of the experts was improved by 5.98%, 5.20%, 4.82%, and 8.19% using our model as an AI-assisted tool. The consistency rates and Cohen's kappa coefficients between experts' interpretations of microscopic digital images and microscopic examination under eyepiece exceeded 93% and 0.83 for the three different morphologic states of yeast. Our model exhibits superior diagnostic accuracy for VVC compared to experts. Experts can significantly improve their own diagnostic accuracies by using our model as an AI-assisted tool. The microscope images collected from each slide effectively represent the slide.

Importance: A cascaded deep neural network model was developed for slide-level diagnosis of vulvovaginal candidiasis (VVC), demonstrating superior diagnostic accuracy compared to experts. Experts significantly enhanced their diagnostic accuracies by utilizing our model as an AI-assisted tool. Therefore, this model holds potential for clinical application to aid in the diagnosis of VVC.