Cephalometric landmark detection using vision transformers with direct coordinate prediction.

Area of Science:

• Medical Imaging
• Computer Vision
• Orthodontics

Background:

• Cephalometric Landmark Detection (CLD) is essential for orthodontic diagnosis and treatment planning.
• Current Deep Learning methods, particularly Convolutional Neural Networks (CNNs) using heatmap prediction, exhibit insufficient performance for large-scale clinical use.

Purpose of the Study:

• To introduce a novel approach for CLD using Vision Transformers (ViTs) with direct coordinate prediction.
• To evaluate the performance of ViTs against contemporary CNN architectures and heatmap-based methods.

Main Methods:

• Utilized Vision Transformers (ViTs) for direct coordinate prediction, bypassing memory-intensive heatmap generation.
• Conducted extensive ablation studies comparing ViTs against ConvNeXt V2 (CNN) and Segformer (heatmap-based).

Main Results:

• ViTs with direct coordinate prediction achieved superior performance, exceeding state-of-the-art CLD methods by over 2 mm in mean radial error.
• Non-adapted CNNs performed poorly, while existing methods showed limited generalization, especially on sparse datasets.

Conclusions:

• Vision Transformers with direct coordinate prediction demonstrate significant promise for advancing CLD and medical computer vision.
• The proposed method offers improved accuracy and generalization capabilities for clinical orthodontic applications.