Oral-Anatomical Knowledge-Informed Semi-Supervised Learning for 3D Dental CBCT Segmentation and Lesion Detection

Area of Science:

• Dental Imaging and AI
• Medical Image Analysis
• Machine Learning in Healthcare

Background:

• Cone beam computed tomography (CBCT) is crucial in dental healthcare for diagnosis and treatment planning.
• Manual segmentation of 3D CBCT images is time-consuming and requires specialized expertise.
• Automating segmentation with AI faces challenges due to the need for large, labeled datasets.

Purpose of the Study:

• To develop an AI model for automated 3D CBCT image segmentation and lesion detection.
• To address the limitation of data dependency in AI models for dental imaging.
• To improve the efficiency and accuracy of lesion detection in early-stage dental conditions.

Main Methods:

• Proposed a novel Oral-Anatomical Knowledge-informed Semi-Supervised Learning (OAK-SSL) model.
• Integrated qualitative oral-anatomical knowledge into a deep learning framework.
• Developed knowledge-informed dual-task learning and a semi-supervised loss function.

Main Results:

• OAK-SSL demonstrated superior performance in segmenting 3D CBCT images compared to existing methods.
• The model effectively segmented small lesions, which are clinically significant for early treatment.
• Achieved robust, accurate, and generalizable segmentation results on a real-world dataset.

Conclusions:

• OAK-SSL offers a promising approach to automate 3D CBCT segmentation and lesion detection.
• Integrating domain knowledge significantly enhances AI model performance in dental imaging.
• This AI-driven method can improve diagnostic accuracy and treatment planning in dentistry.