YOLOv8m-segmentation for detecting cervical burnout and caries in bitewing radiographs:A deep learning approach
- Author:
Mohd Isyrafuddin Bin ISMAIL
1
;
Nooritawati Md TAHIR
;
Wan Syahirah Binti W. SAMSUDIN
;
Mas Suryalis AHMAD
;
Nashuha OMAR
;
Mohd Yusmiaidil Putera Mohd YUSOF
Author Information
- Publication Type:Original Articles
- From:Imaging Science in Dentistry 2026;56(1):26-35
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:This study evaluated the performance of the YOLOv8m-seg model in detecting and delineating interproximal caries and cervical burnout on bitewing radiographs and examined whether increasing the number of training epochs improved segmentation accuracy and consistency.
Materials and Methods:In total, 1,410 bitewing radiographs were annotated using polygon-based masks by a trained dental clinician. The YOLOv8m-seg model was trained for 50, 100, and 150 epochs on 1,128 images and validated on 282 images using the Ultralytics segmentation framework. Model performance was assessed using precision, recall, and mean average precision at intersection-over-union thresholds of 0.5 and 0.5 to 0.95 (mAP0.5, mAP0.5-0.95) for both bounding box and mask outputs. Additional evaluation was conducted on a non-augmented validation subset.
Results:Extended training duration was associated with improved segmentation performance. The highest mask mAP0.5-0.95 value was 0.828 at epoch 150. Both box-based precision and recall increased with longer training, whereas mask-based evaluation more accurately reflected the model’s ability to delineate the boundaries of caries and cervical burnout. Performance appeared consistent across both classes in the augmented validation split but was reduced in the non-augmented validation subset.
Conclusion:The YOLOv8m-seg model demonstrated high diagnostic accuracy in distinguishing proximal caries from cervical burnout on bitewing radiographs. Its mask-based outputs may assist clinicians in early lesion recognition and support improved diagnostic decision-making. Future studies should evaluate model generalizability across broader populations and diverse clinical environments and should prioritize assessment using non-augmented validation sets and independent test datasets.
