Leveraging Pretrained Transformers for Efficient Segmentation and Lesion Detection in Cone-Beam Computed Tomography...

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Cone-beam computed tomography (CBCT) is crucial for detecting jaw lesions but interpretation is challenging and time-consuming.
• Artificial intelligence (AI) in CBCT segmentation can enhance lesion detection accuracy.
• Consistent automated lesion detection is difficult, particularly with limited training data.

Purpose of the Study:

• To evaluate the effectiveness of pretrained transformer-based architectures for semantic segmentation of CBCT volumes.
• To assess the application of these models in periapical lesion detection.

Main Methods:

• CBCT volumes (n=138) were annotated with labels including 'lesion,' 'restorative material,' 'bone,' 'tooth structure,' and 'background.'
• U-Net and Swin-UNETR models (pretrained and from scratch) were trained using subsets of the annotated CBCTs.
• Performance was evaluated using Sørensen-Dice coefficient (DICE) for segmentation and sensitivity/specificity for lesion detection, with varying training sample sizes (20, 40, 60, 103).

Main Results:

• The pretrained Swin-UNETR model achieved high DICE scores across all labels, including 0.8512 for 'lesion' with 103 samples.
• Lesion detection performance remained statistically similar between models trained with 103 and 60 images, with 1.00 sensitivity and 0.94 specificity using 60 images.
• Pretrained Swin-UNETR outperformed Swin-UNETR-SCRATCH and U-Net in segmentation and lesion detection specificity, especially with limited training data.

Conclusions:

• Transformer-based Swin-UNETR architectures demonstrate excellent semantic segmentation and periapical lesion detection capabilities.
• Using pretrained models offers a viable alternative to traditional U-Net architectures, requiring smaller training datasets for effective CBCT analysis.