A novel difficult-to-segment samples focusing network for oral CBCT image segmentation.

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of oral cone-beam computed tomography (CBCT) images is crucial for clinical dentistry.
• Current deep learning methods struggle with blurred contours and scale variations, leading to poor segmentation of tooth crowns and roots.
• These limitations hinder precise clinical diagnosis and treatment planning.

Purpose of the Study:

• To develop an advanced deep learning model for accurate oral CBCT image segmentation.
• To overcome the limitations of existing methods in segmenting difficult regions like tooth crowns and roots.
• To enhance the clinical applicability of automated segmentation in dentistry.

Main Methods:

• Proposed the Difficult-to-Segment Focus Network (DSFNet), incorporating a Feature Capturing Module (FCM) for robust feature extraction.
• Implemented a Multi-Scale Feature Fusion Module (MFFM) to integrate information across different scales.
• Introduced a hybrid loss function combining Focal Loss and Dice Loss to prioritize difficult-to-segment samples.

Main Results:

• DSFNet achieved a Dice Similarity Coefficient (DSC) of 91.85% and an Average Symmetric Surface Distance (ASSD) of 0.216 mm.
• The model demonstrated superior performance compared to existing dental CBCT segmentation techniques.
• Experimental results confirm the effectiveness in segmenting challenging oral CBCT regions.

Conclusions:

• DSFNet offers a significant advancement in oral CBCT image segmentation accuracy.
• The proposed network effectively handles blurred contours and scale differences, improving segmentation of critical dental structures.
• DSFNet shows strong potential for real-world clinical application in dentistry.