YOLO-guided LABNet: advanced deep learning model with color-attention for light-independent automatic tooth color prediction
Neural Computing and Applications, cilt.38, sa.16, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 38 Sayı: 16
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s00521-026-12376-6
- Dergi Adı: Neural Computing and Applications
- Derginin Tarandığı İndeksler: Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, Index Islamicus, INSPEC, zbMATH, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Image processing, Squeeze-and-excitation, Tooth color prediction, YOLOv5
- Atatürk Üniversitesi Adresli: Evet
Özet
Accurate tooth color determination is important for successful treatment outcomes in restorative and aesthetic applications in dentistry. Tooth color determination in prosthetic dentistry remains a challenging process due to variability in light sources, subjectivity in visual assessment, the high cost of instrumental devices, and lack of standardization. In this study, we proposed YOLO-Guided LABNet that is enhanced CNN model with color-attention for light-independent automatic tooth color prediction. A new dataset of patient intraoral dental images obtained under different clinical lighting sources is used in this study. YOLOv5 was used for tooth localization and the proposed problem-specific LABNet model was used for color classification. The YOLOv5 model was used for automatic detection of tooth regions. Teeth were transformed into LAB color space. Contrast brightness balance was achieved with CLAHE applied to the L canal. Tooth ROIs were identified and the dataset was made suitable for classification. The LABNet model has been proposed as a color classification model robust to light variability using Static Color Channel Scaling (SCCS) and Dynamic Color Channel Attention (DCCA) blocks. In this study, YOLOv5 achieved 99.5% mAP@50 in tooth detection, while the LABNet model achieved 89% accuracy in the validation phase compared to pre-trained models such as VGG16, VGG19, EfficientNetB0 and ResNet50. This study presents a scalable system that automates, standardizes and improves reproducibility of digital tooth shade prediction in real clinical conditions, providing an objective and reliable solution for clinical applications.