Intra-YOLO: A Small Object Detection Model for Caries and Molar-Incisor Hypomineralization in Intraoral Photography Based on Transfer Learning with Reinforcement Learning

Technical Summary: Intra-YOLO for Small Object Detection in Intraoral Photography

Problem Statement
The paper addresses the clinical challenge of differentiating between dental caries and Molar-Incisor Hypomineralization (MIH) using intraoral photography. While both conditions present similarly, their etiologies and treatments differ significantly; MIH is a developmental defect, whereas caries results from bacterial demineralization. Accurate diagnosis is hindered by the small size of lesions (over 78% of annotated lesions occupy less than 0.58% of the image area), variability in imaging conditions (e.g., saliva contamination, lighting), and the subjective nature of visual examinations by clinicians. Existing research has largely focused on radiographic imaging or single-tooth classification, leaving a gap in automated, full-mouth detection systems capable of handling small, scale-variant lesions in complex intraoral environments.

Methodology
The authors propose Intra-YOLO, a deep learning framework optimized for small object detection, built upon the YOLOX architecture. The methodology integrates several advanced techniques:

1. Teacher-Student Paradigm with Transfer Learning:

   • Teacher Model (S-YOLO): Trained on high-resolution cropped image patches (640×640 pixels) to maximize feature extraction from small lesions. This model incorporates:
     • SSM (State-Space Modeling) Attention Mechanisms: To suppress noise and background interference.
     • S-PAFPN (Enhanced Feature Pyramid Network): Combines high-resolution details from large feature maps with semantic information from early backbone layers to preserve fine-grained features.
     • Sliced Inference: Applied to handle the small size of lesions effectively.
   • Student Model: Trained on full-size intraoral images to integrate both localized and global features.

2. Reinforcement Learning (RL) for Knowledge Distillation:

   • Instead of static distillation, the framework employs an RL agent (using Proximal Policy Optimization, PPO) to dynamically decide which predicted bounding boxes from the teacher should be used as pseudo-labels for the student.
   • The agent's decision is based on a reward function considering lesion size, the teacher's confidence score, and the student's prediction uncertainty (measured by entropy loss). This ensures the student focuses on "high-value" learning cases, particularly for small and variable-sized lesions.

3. Data Processing:

   • The study utilized a clinical dataset of 822 intraoral images from Taipei Municipal Wanfang Hospital, containing 3,583 caries and 4,443 MIH annotations.
   • The training set was augmented by cropping images into patches, resulting in 6,417 cropped images with over 12,000 bounding boxes to mitigate background imbalance and increase valid training samples.

Key Contributions

• Intra-YOLO Framework: A novel architecture specifically designed for the dual detection of caries and MIH in intraoral photography, addressing the specific challenge of extremely small lesion sizes.
• Scale-Aware Distillation: The integration of reinforcement learning into the knowledge distillation process allows for adaptive selection of training samples, improving the model's ability to learn from difficult, small-scale targets without introducing noisy supervision.
• Enhanced Feature Fusion: The development of the S-PAFPN module and the application of SSM attention mechanisms specifically tailored for intraoral imagery to enhance multi-scale feature fusion and noise suppression.

Results
The model was evaluated on a held-out clinical dataset using mean Average Precision (mAP) and mAP50 (at IoU 0.5).

• Performance: Intra-YOLO achieved an mAP of 23.1% and an mAP50 of 47.8%.
• Comparisons:
  • Outperformed the baseline S-YOLO (without RL distillation), which achieved 22.0% mAP and 44.2% mAP50.
  • Significantly outperformed other end-to-end models, including two-stage detectors (Faster R-CNN, DETR, DAB-DETR) which ranged from 12.4% to 16.7% mAP, and single-stage detectors like YOLOv5 and standard YOLOX with sliced inference (19.7% mAP).
• Ablation: Experiments confirmed that the PPO-based adaptive distillation strategy yielded the best results compared to static thresholding strategies, with specific improvements in small object detection (mAPs reaching 18.3%).

Significance and Claims
The paper claims that Intra-YOLO provides a computer-aided diagnosis (CAD) system capable of accurately and efficiently detecting caries and MIH in intraoral photographs. The authors posit that this system can:

• Reduce diagnostic variability and subjectivity among clinicians.
• Support timely, evidence-based treatment planning by providing objective references for lesion localization and classification.
• Facilitate earlier clinical decision-making.

The authors maintain a modest tone, acknowledging that while the results are promising, further validation on larger and more diverse datasets is required before the system can be fully integrated into routine dental practice. They emphasize that the system is designed to support clinicians rather than replace them, specifically targeting the reduction of misdiagnosis between visually similar lesions.