Temporomandibular joint CBCT image segmentation via multi-view ensemble learning network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of the temporomandibular joint (TMJ) in cone beam CT (CBCT) images is crucial for diagnosing temporomandibular joint osteoarthrosis (TMJOA).
• Existing 3D medical image segmentation methods, often based on convolutional neural networks, require significant GPU memory and computational power due to the need for extensive global context and spatial information.

Purpose of the Study:

• To develop a novel network, MVEL-Net (Multi-view Ensemble Learning Network), for efficient and accurate segmentation of TMJ from CBCT images.
• To address the computational challenges associated with 3D medical image segmentation by proposing a method that utilizes multiple views and ensemble learning.

Main Methods:

• The MVEL-Net employs resampling of 3D CBCT images along three dimensions to create multiple 'weak learners,' each capturing different spatial semantic information.
• A subsequent 'strong learning' network integrates the outputs from these weak learners to achieve enhanced segmentation accuracy.
• The model was evaluated on a clinical dataset of 88 subjects' TMJ CBCT images.

Main Results:

• The MVEL-Net achieved a high average Dice similarity coefficient (DSC) of 0.9817 ± 0.0049.
• Excellent performance was also demonstrated by an average surface distance of 0.0540 ± 0.0179 mm and a 95% Hausdorff distance of 0.1743 ± 0.0550 mm.
• The network exhibited superior segmentation accuracy for TMJ in CBCT images while consuming less GPU memory compared to other 3D segmentation networks.

Conclusions:

• The proposed MVEL-Net effectively segments TMJ from CBCT images with high accuracy and reduced computational resource requirements.
• The network's ability to capture spatial context makes it a promising approach for various 3D medical image segmentation tasks.
• This method has the potential to facilitate the broader adoption of AI-driven automated analysis for volumetric medical scans.