Improving AI-Based Dental Caries Detection with Limited Data: A Few-Shot Learning Approach
2025 16th International Conference on Electrical and Electronics Engineering, ELECO 2025, İstanbul, Türkiye, 27 - 29 Kasım 2025, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/eleco69582.2025.11329224
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: Caries detection, Dental X-ray, Few-shot learning, Mask R-CNN, Multi-class segmentation
- Ankara Üniversitesi Adresli: Evet
Özet
AI is crucial in medical image analysis, particularly in dentistry, where decision support systems improve diagnostics. However, AI systems often lack labeled data. This study explores few-shot learning in training a Mask R-CNN network for multi-class segmentation of dental X-rays, focusing on caries detection. A dataset of 250 bitewing X-ray samples was augmented to 1253 samples and split into training, validation, and testing sets. The Mask R-CNN network, trained with 1061 samples, achieved 95.03% accuracy in detecting filling-type and root canal-type caries. The network was trained using 10-shot learning with 166 samples, divided into 10 seeds with 5 samples each of caries-1 and caries-2 classes. The average accuracy across seeds for 10-shot learning was 91.46%. Results show few-shot learning effectively distinguishes new caries classes when multiclass bitewing data is limited. This approach offers a solution for developing AI decision support systems in dentistry when facing data scarcity and computational challenges.