Artifact-Aware Fungal Detection in Dermatophytosis: A Transformer-Based Approach for KOH Microscopy

Area of Science:

• Medical Mycology
• Artificial Intelligence in Medicine
• Image Analysis

Background:

• Potassium hydroxide (KOH) microscopy is standard for dermatophytosis diagnosis but faces challenges with artifacts and inter-observer variability.
• Accurate identification of fungal hyphae is crucial for timely and effective treatment.

Purpose of the Study:

• To develop and evaluate a transformer-based object detection framework (RT-DETR) for precise fungal structure localization in KOH microscopy images.
• To benchmark the RT-DETR model against convolutional neural network (CNN) based detectors.

Main Methods:

• A dataset of 2540 KOH microscopy images was manually annotated to distinguish fungal elements from artifacts.
• The RT-DETR model was trained and validated using five-fold stratified cross-validation.
• Performance was compared against YOLOv11 and Faster R-CNN under identical conditions.

Main Results:

• RT-DETR achieved a mean Average Precision (AP@0.50) of 89.73% ± 1.48%, with high precision (0.921 ± 0.014) and recall (0.831 ± 0.011).
• The model demonstrated high image-level sensitivity (0.989 ± 0.022) with minimal missed cases.
• RT-DETR outperformed CNN-based detectors in accuracy and artifact suppression.

Conclusions:

• A transformer-based AI system shows technical feasibility for detecting fungal regions in KOH microscopy.
• This AI system can serve as a valuable decision-support tool for clinicians.
• Further multi-center validation is needed to confirm clinical generalizability.