# Fungal recognition in vaginal discharge using deep learning analysis of mobile device-acquired microscopic images

## TL;DR

This study develops AI models to detect fungal infections in vaginal discharge using images from smartphones and tablets, enabling faster and more accurate diagnosis.

## Contribution

The study introduces deep learning models for automated detection and segmentation of fungal morphologies in mobile device-acquired microscopic images.

## Key findings

- ResNet18 achieved high accuracy in classifying fungal presence with an F1-score of 0.986 and AUC of 0.99.
- YOLOv11 effectively segmented Candida morphologies with strong precision and recall metrics.
- Expert feedback rated the AI results highly, with 4.25/5 on average and only 3.68% of images deemed inappropriate.

## Abstract

Vulvovaginal candidiasis (VVC) is a common fungal infection that is frequently diagnosed through manual microscopic examination of vaginal discharge. Artificial Intelligence (AI)-assisted analysis of microscopic images enables rapid and accurate diagnosis, supporting timely and effective antifungal therapeutic interventions. However, conventional light microscopy often lacks cameras, limiting digital image analysis and AI applications. While mobile devices offer a practical alternative, no AI tools currently exist for the automated detection of fungal cellular morphology in microscopic images captured by smartphones and tablets. In this study, we developed deep learning models to segment fungal morphologies in microscopic images of vaginal discharge acquired with smartphones and tablets.

Three models were developed: ResNet18 for binary classification (Candida presence/absence), YOLOv5 for detection, and YOLOv11 for segmentation. Models were trained using 1,259 microscopy images of Gram-stained vaginal discharge acquired with smartphones or tablets, along with 67 images obtained from conventional microscopes. These images were divided into training, validation, and test sets. Annotated microscopic images for fungal elements were used to train YOLO models in a two-stage approach: Stage 1 utilized 687 annotated images of yeast infections to learn general fungal morphology, comprising 266 bounding box–annotated images sourced from Roboflow and 421 segmentation-labeled images manually annotated from the open-access dataset. Stage 2 fine-tuned the models on the annotated mobile device-acquired dataset. Metrics included F1-score, area under the curve (AUC), precision, recall, and mean average precision at 50% intersection over union (mAP50). Experts assessed segmentation outputs for diagnostic utility, providing explainability to the AI results.

ResNet18 achieved F1-score=0.986, AUC = 0.99. YOLOv5 performed best at IoU=0.50 (precision=0.812, recall=0.622, mAP50 = 0.730); YOLOv11 at IoU=0.25 (precision=0.766, recall=0.700, mAP50 = 0.727). Expert ratings averaged 4.25/5. Only 3.68% of images were rated as inappropriate due to false negative or false positive segmentations.

ResNet18 accurately classified microscopic images for fungal elements, while the YOLOv11 model effectively delineated Candida morphologies, including yeasts, budding yeasts, and filamentous forms from clinical specimens. The high accuracy and positive expert feedback demonstrate the feasibility of integrating AI-assisted mobile microscopy into routine workflows, thereby advancing digital analysis of microbial infections using conventional light microscopy. With further clinical validation and expansion to include other infections, this approach holds great potential to establish robust real-world utility.