OralSegNet: An Approach to Early Detection of Oral Disease Using Transfer Learning.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Oral diseases pose significant health challenges, requiring accurate and timely detection.
• Automated diagnostic tools can improve accessibility and efficiency in oral healthcare.

Purpose of the Study:

• To develop and evaluate a deep learning-based segmentation system for automated detection and localization of oral diseases.
• To investigate the performance of three YOLOv11 architecture variants (YOLOv11n-seg, YOLOv11s-seg, YOLOv11m-seg) on intraoral images.

Main Methods:

• A dataset of intraoral images was curated and augmented through multiple versions (v1-v3) to address class imbalance and enhance generalization.
• Three stages of training were employed: feature extraction, partial fine-tuning, and full fine-tuning.
• The YOLOv11m-seg model demonstrated superior performance in the partial fine-tuning phase.

Main Results:

• The YOLOv11m-seg model achieved a box mAP@50 of 0.521 and a mask mAP@50 of 0.500 during partial fine-tuning.
• Training was conducted on Google Colab's free tier, utilizing available computational resources.
• The best-performing model was successfully exported to ONNX format with NMS enabled.

Conclusions:

• The developed deep learning system effectively detects and localizes oral diseases from intraoral images.
• The system was deployed as a client-side web application using React.js and ONNX Runtime Web for broad accessibility.
• This tool empowers both clinicians and non-experts to perform rapid oral disease detection with a single click.