V²-Net: An Attention-guided Volumetric Regression Network for Tooth Landmark Localization on CT Images with Metal...

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Artificial Intelligence

Background:

• Manual tooth landmark localization on CT images is crucial for virtual surgical planning in orthognathic surgery.
• This manual process is time-consuming, labor-intensive, and requires specialized expertise.
• Automated localization is challenging due to low resolution and metal artifacts in dental CT images.

Purpose of the Study:

• To develop an attention-guided volumetric regression network (V²-Net) for accurate tooth landmark localization on low-resolution CT images with metal artifacts.
• To enhance the precision of tooth landmark identification for improved virtual surgical planning.

Main Methods:

• Proposed an attention-guided volumetric regression network (V²-Net) with a coarse-to-fine-attention mechanism.
• Integrated attention-guided learning and a 3D attention module with optimal Pseudo Huber loss.
• Evaluated the network's performance on CT images with inherent resolution and artifact challenges.

Main Results:

• Achieved state-of-the-art accuracy with a mean radial error of 0.85 ± 0.40 mm.
• Demonstrated significant improvements in localization accuracy due to attention-guided learning and the 3D attention module.
• Attained a 97.92% success detection rate within the clinically acceptable 2.0 mm accuracy range.

Conclusions:

• The V²-Net effectively addresses challenges of low resolution and metal artifacts in dental CT imaging.
• The proposed method offers a highly accurate and reliable solution for tooth landmark localization in orthognathic surgery planning.
• This advancement has the potential to streamline virtual surgical planning workflows and improve patient outcomes.