Automatic cephalometric landmark detection using deep learning
A Fully Deep Learning Model for the Automatic Identification of Cephalometric Landmarks
Authors: Young Hyun Kim, Chena Lee, Eun-Gyu Ha, Yoon Jeong Choi, Sang-Sun Han
Summary
This study from Yonsei University College of Dentistry (Seoul, Korea) presents a fully automatic deep learning model for detecting 13 key cephalometric landmarks on lateral cephalometric radiographs.
A total of 950 images were collected from clinical archives at Yonsei Dental Hospital, annotated by two orthodontic experts (15 and 5 years of experience) after calibration sessions.
The proposed framework consists of two CNN-based modules:
- ROI Machine – identifies 13 regions of interest containing each landmark.
- Detection Machine – performs precise coordinate prediction for each ROI using CNNs (8 convolution + 5 pooling + 2 fully connected layers, ELU activation).
The network was trained on 800 images, validated on 100, and tested on 50 using Ubuntu 18.04 / NVIDIA Titan Xp GPU (12 GB).
Performance was evaluated with Mean Radial Error (MRE), Successful Detection Rate (SDR), and inter-examiner variability.
Results:
- Average clinically acceptable range between experts = 1.24 mm
- Model MRE: 1.84 mm (lower than human variability for A-point, UIB, LIB, ANS)
- Successful Detection Rate: 36.1 % within expert variability
- Inter-examiner agreement: ICC > 0.99 for all 13 landmarks
These results demonstrate that the proposed model achieved accuracy comparable to expert performance and even surpassed trained clinicians for certain landmarks.
DOI: https://doi.org/10.5624/isd.20210077
Key Words
Cephalometric Radiography, Anatomic Landmarks, Automatic Landmark Detection, Deep Learning, Neural Network Models, CNN, Orthodontics, Computer Vision, Dental AI, Radiographic Analysis
Extracted Data
- Year: 2021
- Modality: Lateral Cephalometric Radiography
- Dataset: 950 images (800 train / 100 validation / 50 test)
- Dataset Split: 84 % train | 11 % validation | 5 % test
- Landmarks Detected: 13 hard-tissue points (S, N, Or, Po, A, B, Pog, Me, UIB, LIB, PNS, ANS, Ar)
- Network Architecture: Custom CNN (two-step ROI + Detection Machine, 8 Conv + 5 Pool + 2 FC)
- Metrics: MRE = 1.84 mm | ICC > 0.99 | SDR = 36.1 % (expert range)
- AP – Strategy: Manual annotation by two orthodontists + ROI-based deep learning with clinically acceptable error defined by expert variability
- AP – Professional Qty: 2 orthodontists
- AP – Supervisor Presence: No Information
- AP – Experience Level: 15 and 5 years in Orthodontics
- AP – Expertise Area: Orthodontics
- AP – Tool or System: Custom CNN implemented in Python (TensorFlow / Keras)
- Task: Landmark Detection / Point Localization
- Project Objective: To develop a fully automatic deep learning framework for cephalometric landmark identification using clinical data and evaluate it against expert variability.
Clinical Relevance
- Clinical importance: Cephalometric landmark identification is crucial for orthodontic diagnosis but is time-consuming and subject to human error.
- Innovation: The model introduces a two-stage deep learning structure that mimics clinical workflow (ROI + detection), reducing inter-observer variability.
- Practical impact: Fully automated landmark detection can enhance efficiency in orthodontic planning, research datasets, and AI-assisted cephalometric analysis.
DOI: https://youtu.be/uK7OSfyoSjc
