What question did this study set out to answer?

This study aims to develop a self-supervised deep learning model for the identification of clinically relevant Candida species from microscopic images.

July 5, 2026Scientific Reports0 citationsOpen Access

Automated identification of clinically important Candida yeast species for microscopic images using self-supervised learning

FMFingani Annie Mphande-NyasuluKing Mongkut's Institute of Technology Ladkrabang PTPrasert TrivijitsilpKing Mongkut Memorial Hospital DMDr Suchanun MeksangKing Mongkut's University of Technology North Bangkok

Key Points

This study aims to develop a self-supervised deep learning model for the identification of clinically relevant Candida species from microscopic images.
Developed and validated a self-supervised deep learning model (DINOv2).
Utilized YOLOv4 tiny for detection and UNet for segmentation of microscopic images.
Performed binary and four-class classification with cross-validation.
Achieved a mean average precision (mAP@50) of 0.908 and AUC of 0.949.
Binary classification accuracies ranged from 0.980 to 0.988 with F1 score of 0.981.
Four-class classification yielded highest recall, precision, and F1 score of 0.977.

Abstract

Opportunistic fungal infections are an escalating global challenge, impacting over a billion people worldwide, particularly immunocompromised individuals. Effective management strategies are urgently needed. While laboratory culture remains the presumptive identification of fungal infections, it is time-consuming, requires expensive equipment, and suffers from observer-dependent variability. Innovative automated tools integrating artificial intelligence (AI) offer promising solutions for timely diagnosis and improved treatment outcomes. This study aims to develop and validate a self-supervised deep learning (SSL) model (DINOv2) to automatically identify clinically relevant Candida species from microscopic images. To begin with, appropriate data preparation, combining detection and segmentation techniques were conducted. Microscopic images were processed using the YOLOv4 tiny model to locate the organisms of interest, followed by image segmentation using the UNet algorithm. Study outcomes showed that YOLOv4 tiny and SSL DINOv2 detection models demonstrated high performance, achieving a mean average precision (mAP@50) of 0.908 and an AUC under the precision-recall (PR) curve of 0.949. Additionally, the semantic segmentation model yielded a Dice score of 0.930, an Intersection over Union (IOU) score of 0.874, and a Hausdorff Distance score of 67.101, confirming the reliability of the dataset for further analysis. Binary classification of two fungal species, Candida albicans and C. krusei , was performed. The SSL models achieved accuracies ranging from 0.980 to 0.988, outperforming baseline models such as Vision Transformer. Similarly, the SSL models achieved an F1 score of 0.981, significantly higher than the baseline models’ score of 0.938. For four-class classification, the small version of DINOv2 trained on 224 × 224-pixel images achieved the highest recall, precision, and F1 score values of 0.977, demonstrating superior performance. Additionally, five-fold cross-validation (accuracy: 94.8–98.3%; AUC-ROC: 0.990–0.998) and an almost perfect level of agreement analysis (κ = 0.845) demonstrate robust potential as a proof-of-concept for clinical applications, pending multi-center validation. In conclusion, the hybrid models exhibited exceptional performance, highlighting their potential as a research prototype for healthcare applications, requiring rigorous prospective validation. However, further validation with diverse datasets is essential to enhance and confirm the feasibility of deploying AI models in the healthcare sector.

Ask AI

Helpful

Bookmark

View Full Paper