Artefact-Aware Fungal Detection in Dermatophytosis: A Real-Time Transformer-Based Approach for KOH Microscopy

This study introduces a real-time, transformer-based RT-DETR framework that achieves highly accurate, artefact-aware detection of fungal hyphae in KOH microscopy images, demonstrating 100% sensitivity and 98.8% accuracy to serve as a reliable automated screening tool for dermatophytosis.

The Gist

Imagine you are a detective trying to find a very specific, tiny clue (a fungal hypha) hidden inside a messy crime scene (a microscope slide). The problem is, the crime scene is covered in debris, dust, and random scratches that look exactly like the clue you're looking for. If you miss the clue, the patient doesn't get treated. If you mistake a scratch for a clue, you might treat someone who isn't sick.

This paper is about building a super-smart, tireless digital detective that can look at these messy microscope slides and find the fungal clues instantly and accurately, even when they are hiding among the noise.

Here is the breakdown of how they did it, using simple analogies:

1. The Problem: The "Needle in a Haystack"

Doctors currently use a method called KOH microscopy to check for fungal infections (like athlete's foot or ringworm). They dissolve the dead skin cells to make the fungus visible.

• The Issue: Even after dissolving the skin, the slide is still full of "junk"—air bubbles, fibers, and skin debris. These "junk" items often look exactly like the fungus.
• The Human Limit: Human doctors get tired. Their eyes can get confused by the junk, leading to missed diagnoses or false alarms. It's like trying to find a specific type of leaf in a pile of leaves, twigs, and plastic wrappers while wearing foggy glasses.

2. The Solution: A "Smart Search Engine" for Microscopes

The researchers built an Artificial Intelligence (AI) system using a technology called RT-DETR.

• The Analogy: Think of old AI models as a person looking at a photo and guessing, "Is there a dog here?" (Yes/No).
• The New AI: This new model is like a laser-guided search engine. Instead of just saying "Yes, fungus is here," it draws a box around the exact spot where the fungus is.
• The Secret Weapon (The Transformer): Most AI looks at small pieces of the image one by one. This AI (a "Transformer") looks at the whole picture at once. It understands the context. It knows, "This looks like a fungus, but it's sitting right next to a giant air bubble, so it's probably just a bubble." It uses "global attention" to see the big picture, not just the details.

3. The Training: Teaching the AI to Spot "Imposters"

To teach the AI, the researchers didn't just show it pictures of fungus. They showed it pictures of fake fungus (artifacts) too.

• The Strategy: They labeled the real fungus with Green Boxes and the confusing junk (fibers, bubbles) with Purple Boxes.
• The Lesson: They told the AI: "Don't just look for the shape of the fungus. Learn to tell the difference between a real fungus and a piece of lint that looks like a fungus."
• The Result: The AI learned to ignore the "imposters" and focus only on the real suspects.

4. The Results: The Perfect Safety Net

They tested this AI on over 2,500 microscope slides.

• Object Level (Finding the Clue): The AI found 97% of the actual fungus pieces. It made a few mistakes where it thought a piece of lint was fungus (about 20% of the time), but that's okay because we can double-check those.
• Image Level (The Diagnosis): This is the most important part. When the AI looked at the whole slide to decide "Is this patient infected or not?", it got 100% right.
  • It never missed a single infected patient.
  • It correctly identified almost all healthy patients.
• The Metaphor: Imagine a security guard at an airport. If the guard misses one bad guy, that's a disaster. If the guard stops a few innocent people for extra checks, that's annoying but safe. This AI is the ultimate security guard: It never lets a bad guy (infection) slip through the door.

5. Why This Matters

• Speed: The AI can analyze a slide in less than 50 milliseconds (faster than a human can blink).
• Reliability: It doesn't get tired, it doesn't have bad eyesight, and it doesn't get confused by the messy background.
• The Future: This isn't meant to replace the doctor. It's meant to be a super-assistant. It highlights the suspicious spots on the screen for the doctor to look at, making the diagnosis faster and more accurate.

In a Nutshell

The researchers built a digital microscope assistant that is incredibly good at ignoring the "visual noise" (dust, bubbles, scratches) and finding the tiny, hard-to-see fungus. By teaching it to spot the "fakes" as well as the "real thing," they created a system that ensures no infected patient is ever missed, bridging the gap between complex computer science and saving lives in a dermatologist's office.

Technical Summary

1. Problem Statement

Dermatophytosis (fungal infection of the skin/nails) is typically diagnosed via Potassium Hydroxide (KOH) microscopy. However, this method faces significant challenges:

• Visual Ambiguity: KOH preparations often contain residual keratin, air bubbles, debris, and fibers that visually mimic fungal hyphae, leading to high inter-observer variability.
• Limitations of Current AI: Existing deep learning approaches largely rely on image-level classification (infected vs. healthy) or pixel-wise segmentation. These methods often fail to distinguish between true fungal structures and confounding artefacts, leading to false positives or missed diagnoses.
• Lack of Object-Level Detection: There is a scarcity of systems that perform precise, object-level localization (bounding boxes) of fungal elements while explicitly modeling artefacts as a distinct class to reduce diagnostic confusion.

2. Methodology

A. Dataset Construction

• Source: 2,540 high-resolution (2048×2048 pixels) KOH microscopy images acquired from routine clinical practice at Istanbul Research and Training Hospital.
• Annotation Strategy: A multi-class annotation approach was employed by expert clinicians:
  • Positive Class: 631 annotations of fungal elements (hyphae and spore clusters).
  • Negative/Artefact Class: 381 annotations of confounding artefacts (keratin debris, fibers, refractive edges).
  • Significance: By explicitly labeling mimics rather than treating them as background, the model is trained to actively discriminate between true pathogens and visual noise.
• Split: The dataset was stratified into training (80%), validation (10%), and testing (10%) sets.

B. Model Architecture

• Core Model: RT-DETR (Real-Time Detection Transformer), specifically the RT-DETR-L variant pre-trained on COCO.
• Mechanism:
  • Utilizes a hybrid CNN-Transformer encoder-decoder pipeline.
  • Employs attention-driven query selection to enable direct bounding box regression and classification without anchor boxes.
  • Leverages global self-attention to distinguish biological structures from mimics in complex, low-contrast backgrounds.
• Training Configuration:
  • Hardware: NVIDIA RTX 4090 GPU.
  • Preprocessing: Images resized to 1024×1024 via letterbox padding; normalized using ImageNet statistics.
  • Augmentation: Aggressive augmentations (e.g., MixUp) were disabled to preserve the structural integrity of thin hyphae. Standard geometric variations (flipping, scaling, minor rotation) were used.
  • Optimization: AdamW optimizer, cosine-annealed learning rate decay, mixed-precision training, and early stopping (patience = 50).
  • Inference: Real-time processing (<50ms per image) with a confidence threshold of 0.25.

3. Key Contributions

1. Artefact-Aware Framework: First study to explicitly model KOH artefacts as a distinct detection class within a transformer-based object detection pipeline, significantly reducing false positives caused by visual mimics.
2. Object-Level Localization: Moves beyond binary classification to provide precise bounding boxes for fungal elements, offering interpretable visual cues for clinicians.
3. Real-Time Performance: Demonstrates that transformer-based models (RT-DETR) can achieve real-time inference speeds suitable for clinical workflows, overcoming the latency often associated with complex deep learning models.
4. Clinical Safety Focus: The system is designed with "perfect sensitivity" as a priority, ensuring no positive cases are missed, which is critical for screening tools.

4. Results

Object-Level Performance (Detection)

• Recall (Sensitivity): 0.9737 (Captures 97.37% of fungal elements).
• Precision: 0.8043.
• AP@0.50: 93.56%.
• Mean IoU: 0.8560, indicating strong spatial agreement between predicted boxes and ground truth.
• Capability: Successfully localized low-contrast hyphae and separated overlapping structures in dense fields.

Image-Level Performance (Clinical Diagnosis)

• Strategy: A case is classified as positive if at least one fungal element is detected with high confidence.
• Sensitivity: 100% (Correctly identified all 89 positive cases; 0 False Negatives).
• Specificity: 98.18%.
• Overall Accuracy: 98.82%.
• Missed Diagnoses: 0.

Failure Analysis

• False Positives: Primarily caused by synthetic fibers, scratch marks, and sharp keratin edges that closely mimic hyphal morphology.
• False Negatives: Rare; occurred only in cases of extremely faint or out-of-focus hyphal fragments with very low contrast.

5. Significance and Impact

• Bridging the Gap: The study successfully bridges the gap between "black-box" AI predictions and clinical explainability by providing visual localization (bounding boxes) that aligns with how dermatologists scan slides.
• Clinical Utility: The 100% sensitivity at the image level positions this system as a highly reliable automated screening tool. It ensures that no infected patient is overlooked, addressing the critical need for early and accurate diagnosis.
• Workflow Integration: By reducing inter-observer variability and fatigue associated with manual scanning of large microscopic fields, the tool can standardize diagnostics in dermatomycology.
• Future Directions: The authors acknowledge the need for multi-center validation and future work to differentiate specific fungal species and integrate anomaly detection for non-fungal mimics (e.g., subungual melanoma).

In conclusion, this paper presents a robust, real-time AI solution that leverages transformer architecture and multi-class training to solve the specific challenges of artefact-rich KOH microscopy, offering a safe and effective decision-support tool for clinicians.