Craniomaxillofacial landmarks detection in CT scans with limited labeled data via semi-supervised learning

Leran Tao ^1,2,3, Xu Zhang ⁴, Yang Yang ^5,6

⁰Department of Oral and Cranio-Maxillofacial Surgery, Shanghai Ninth People's Hospital, College of Stomatology, Shanghai Jiao Tong University School of Medicine, Shanghai, 200011, China.

Heliyon +

|

August 12, 2024

View abstract on PubMed

Summary

Semi-supervised learning (SSL) models improve 3D cephalometric analysis by accurately detecting landmarks in craniomaxillofacial (CMF) CT scans with less labeled data. This reduces medical professional workload and enhances diagnostic accuracy.

Area Of Science

Medical Imaging
Artificial Intelligence in Healthcare
Computer-Aided Diagnosis

Background

Three-dimensional cephalometric analysis is vital for craniomaxillofacial (CMF) assessment, relying on accurate landmark detection in CMF CT scans.
Training deep learning models for this task demands extensive, expertly annotated CMF CT datasets, which are costly and time-consuming to acquire.
The availability of large unlabeled CMF CT datasets presents an opportunity for semi-supervised learning (SSL) approaches.

Purpose Of The Study

To develop and evaluate a semi-supervised learning (SSL) model for accurate landmark detection in 3D CMF CT scans.
To reduce the reliance on large, manually annotated datasets for training deep learning models in cephalometric analysis.
To enhance the efficiency and accuracy of 3D cephalometric analysis by leveraging limited labeled data with abundant unlabeled data.

Main Methods

Developed CephaloMatch, an SSL model utilizing a strong-weak perturbation consistency framework.
Incorporated head position rectification via coarse detection to improve consistency between labeled and unlabeled data.
Employed a multilayers perturbation method to expand the perturbation space for enhanced model robustness. Assessed on 362 CMF CT scans (60 labeled, 288 unlabeled).

Main Results

The SSL model achieved a detection error of 1.60 ± 0.87 mm, outperforming a fully supervised model (1.94 ± 1.12 mm).
The SSL model attained comparable accuracy (1.91 ± 1.00 mm) using only half the labeled data compared to the fully supervised approach.
Demonstrated significant improvement over conventional fully supervised learning methods in landmark detection accuracy.

Conclusions

The proposed SSL model effectively performs landmark detection in 3D CMF CT scans with limited labeled data.
This approach significantly reduces the annotation workload for medical professionals.
The study highlights the potential of SSL to enhance the accuracy and efficiency of 3D cephalometric analysis.

Keywords:

3D cephalometric analysis Computer-assisted surgery design Dentomaxillofacial deformities Landmarks detection Semi-supervised learning