Comparative Analysis of Deep Learning-Based Segmentation Models for Root Canal Filling Detection with Efficiency and Pruning Evaluation


Genc M. Z., ÇELİK B., ÇELİK M. E.

18th International Conference on Electronics, Computers and Artificial Intelligence, ECAI 2026, Bucharest, Romanya, 2 - 03 Temmuz 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/ecai69016.2026.11613830
  • Basıldığı Şehir: Bucharest
  • Basıldığı Ülke: Romanya
  • Anahtar Kelimeler: computational cost, deep learning, dentistry, pruning
  • Gazi Üniversitesi Adresli: Evet

Özet

Accurate automated segmentation of root canal fillings in periapical radiographs represents a critical bottleneck in scaling endodontic quality assessment, yet most existing deep learning approaches prioritize accuracy while neglecting deployment feasibility in resource-constrained clinical settings. This study evaluates deep learning-based segmentation models for automated segmentation of root canal fillings in periapical dental radiographs. Five widely used architectures - U-Net, U-Net++, Feature Pyramid Network (FPN), LinkNet, and SegFormer - were compared in terms of both segmentation performance and computational efficiency using a dataset of 597 annotated periapical radiograph images. Experimental results demonstrated that U-Net achieved the highest segmentation accuracy with a mean Intersection over Union (IoU) of 74.57 and a Dice score of 85.42 across five independent training cycles. Lightweight models such as FPN and LinkNet provided a favorable balance between performance and efficiency for resource-constrained clinical environments, with LinkNet achieving a GPU latency of 7.52 ms and throughput of 132.9 images/s. Additionally, pruning experiments on the best-performing U-Net model revealed that moderate unstructured pruning led to limited performance degradation recoverable through fine-tuning, whereas aggressive pruning and structured pruning at higher ratios resulted in substantial accuracy loss. These results highlight the trade-off between segmentation accuracy and computational efficiency and provide practical insights for the deployment of deep learning models in clinical dental imaging workflows.