97% mAP in FDI-based tooth segmentation on bitewing X-rays using Mask R-CNN

An Enhanced Tooth Segmentation and Numbering According to FDI Notation in Bitewing Radiographs

Authors: Buse Yaren Tekin, Caner Ozcan, Adem Pekince, Yasin Yasa

---

Summary This study introduces a deep learning–based method for automatic tooth segmentation and numbering in bitewing radiographs, applying the Mask R-CNN architecture with FDI numbering notation.
The dataset comprised 1,200 bitewing radiographs (1,000 train / 200 test) acquired at Ordu University. Images were annotated by oral and maxillofacial radiologists using polygon masks for each tooth region (crowns + roots).

The Mask R-CNN model (ResNet-101 + FPN backbone) was trained for 400 epochs with Adam optimizer (lr = 0.001) using transfer learning from MS COCO weights.
The system achieved:

• Segmentation: Precision = 100 %, mAP = 97.49 %, Recall = 97.24 %, F1 = 97.36 %
• Numbering: Precision = 94.35 %, mAP = 91.51 %, Recall = 95.20 %, F1 = 93.33 %

Comparative analysis with 12 other CNN architectures (AlexNet, VGG-16, DenseNet, MobileNet-v2, GoogleNet, etc.) confirmed Mask R-CNN as the most accurate approach.
The model automatically assigns FDI tooth numbers (ISO-3950), detects overlapping teeth, and improves numbering accuracy compared with prior methods.

DOI: https://doi.org/10.1016/j.compbiomed.2022.105547

---

Key Words

Bitewing Radiography, Tooth Segmentation, Tooth Numbering, Mask R-CNN, Instance Segmentation, FDI Notation, ResNet-101, Deep Learning, Computer Vision, Dental AI, Oral Radiology

---

Extracted Data

• Year: 2022
• Modality: Dental Bitewing Radiography
• Dataset: 1,200 images (1,000 train / 200 test)
• Dataset Split: Train = 1,000 | Test = 200
• Network Architecture: Mask R-CNN (ResNet-101 + FPN backbone)
• Metrics: Segmentation → Precision 100 % | mAP 97.49 % | F1 97.36 %; Numbering → Precision 94.35 % | mAP 91.51 % | F1 93.33 %
• AP – Strategy: Polygonal annotation by experts + FDI notation + transfer learning from MS COCO
• AP – Professional Qty: 4 Radiologists (Ordu and Karabuk Universities)
• AP – Supervisor Presence: Yes (faculty oversight at Ordu University)
• AP – Experience Level: Experienced clinicians / academic researchers
• AP – Expertise Area: Oral and Maxillofacial Radiology, Computer Engineering
• AP – Tool or System: Custom Mask R-CNN in Python (TensorFlow / Keras / PyTorch)
• Task: Instance Segmentation + Tooth Numbering (FDI Notation)
• Project Objective: To develop a Mask R-CNN framework for accurate tooth segmentation and automatic numbering in bitewing radiographs using FDI notation and transfer learning.

---

Clinical Relevance

• Clinical importance: Bitewing radiographs are essential for detecting proximal caries and restorations but manual tooth identification is time-consuming and prone to error.
• Innovation: The integration of Mask R-CNN with FDI notation enables automatic segmentation and numbering with high precision, validated against expert annotations.
• Practical impact: The framework provides an AI-based workflow to assist radiologists in tooth identification, reduce annotation time, and improve consistency in dental records and forensic applications.

Watch on YouTube: https://youtu.be/Pydca3jAWMU