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

Area of Science:

• Medical Mycology
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Opportunistic fungal infections pose a significant global health challenge, particularly for immunocompromised individuals.
• Traditional laboratory culture for fungal identification is slow, costly, and subjective.
• Automated AI tools are needed for rapid and reliable diagnosis of fungal infections.

Purpose of the Study:

• To develop and validate a self-supervised deep learning (SSL) model (DINOv2) for automatic identification of clinically relevant Candida species from microscopic images.
• To integrate detection and segmentation techniques for precise organism localization and classification.
• To evaluate the performance of the developed AI models against established benchmarks.

Main Methods:

• Data preparation involved combining object detection (YOLOv4 tiny) and image segmentation (UNet).
• Self-supervised deep learning model DINOv2 was employed for image analysis.
• Binary and four-class classification tasks were performed on Candida species, with rigorous cross-validation and agreement analysis.

Main Results:

• The YOLOv4 tiny and SSL DINOv2 detection models achieved high performance (mAP@50=0.908, AUC-PR=0.949).
• Semantic segmentation yielded excellent scores (Dice=0.930, IOU=0.874).
• SSL models demonstrated high accuracy (0.980-0.988) and F1 scores (0.981) for binary classification, outperforming baseline models. Four-class classification achieved an F1 score of 0.977. Cross-validation showed high accuracy (94.8-98.3%) and AUC-ROC (0.990-0.998).

Conclusions:

• The developed hybrid AI models show exceptional performance and potential as a research prototype for healthcare applications in identifying Candida species.
• The study confirms the reliability of the dataset and the robustness of the AI models for clinical applications, pending further validation.
• Rigorous prospective validation with diverse datasets is essential to confirm the feasibility of deploying these AI models in the healthcare sector.