Abstract
OBJECTIVE: Deep learning-based segmentation system is proposed that exploits three variants of YOLOv11 architecture, namely YOLOv11n-seg, YOLOv11s-seg, and YOLOv11m-seg for automated detection and localization of the oral disease conditions from photographic intraoral images. METHOD: Dataset has been created by combining publicly available data from sources like Roboflow resulting in an initial version (v1) having 582 images annotated at the pixel level. To mitigate class imbalance issues as well as increase generalization capability by the model, progressively this dataset was augmented to create version 2 (v2) and then further extended into version 3 (v3). Training was done in three separate stages: Feature extraction, partial fine-tuning, and full fine-tuning. RESULTS: YOLOv11m-seg model performed best in the partial fine-tuning phase with results of box mAP@50 = 0.521 and mask mAP@50 = 0.500. Training was done on Google Colab's free tier using an Intel Xeon CPU with 13 GB RAM, 15 GB T4 GPU, and 120 GB storage allowance. CONCLUSIONS: For application, the best performing model was exported to ONNX format with NMS enabled and deployed as a fully client-side responsive web app built in React.js and ONNX Runtime Web. The tool enables both clinicians and non-experts to detect oral diseases from intraoral images with a single click.