Amsel criteria based computer vision for diagnosing bacterial vaginosis

Abstract

Bacterial vaginosis (BV) is a common vaginal infection that can predispose patients to several complications, such as pelvic inflammatory disease. Like many illnesses, existing diagnostic methods face a trade-off between diagnostic certainty and cost. To help address this dilemma, we explore a computational diagnostic approach implementable as an IoT device. We developed several deep learning models based on the Amsel criteria to evaluate different inexpensive point-of-care tests that better automate the diagnosis of BV. We first determined how to best diagnose BV via computer vision models trained on epithelial cell images. We found that training a ResNet18 model on NuSwab diagnostic labels achieved an 89% F1 score. We then experimented with augmenting computer vision results with other Amsel criteria values through multi-layer perceptrons, finding that also using whiff test values increased performance to an F1 of 91% and to a sensitivity surpassing human-performed Amsel criteria at 94.31%. These results provide the first insight into how combinations of images and other Amsel criteria data can best be used for reliable diagnoses, paving the way for future research into IoT-based BV diagnostics.