CTA-UNet: CNN-transformer architecture UNet for dental CBCT images segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current deep learning models struggle with segmenting dental cone-beam computed tomography (CBCT) images due to complex root morphology and fuzzy boundaries.
• Manual annotation of dental CBCT images is costly and time-consuming, limiting the development of accurate segmentation models.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for segmenting dental CBCT images.
• To address the limitations of existing models in handling complex root features and limited labeled data.

Main Methods:

• Proposed a CNN-Transformer Architecture UNet network combining Convolutional Neural Networks (CNNs) for local feature extraction and Transformers for long-range dependencies.
• Incorporated multiple spatial attention modules to enhance spatial information processing.
• Introduced a novel Masked image modeling method for simultaneous pre-training of CNN and Transformer modules, mitigating small labeled dataset issues.

Main Results:

• The proposed model achieved superior performance with a Dice Similarity Coefficient (DSC) of 87.12% and Intersection over Union (IoU) of 78.90%.
• Achieved Hausdorff Distance 95th percentile (HD95) of 0.525 mm and Average Surface Distance (ASSD) of 0.199 mm, indicating high accuracy.
• Demonstrated effectiveness in segmenting tooth roots and alveolar bone in dental CBCT images.

Conclusions:

• The developed CNN-Transformer UNet offers a more efficient and effective approach for automatic dental CBCT image segmentation.
• The method shows significant potential for real-world applications in orthodontics and dental implantology.
• The novel pre-training strategy effectively addresses challenges posed by limited annotated data.