PXseg: automatic tooth segmentation, numbering and abnormal morphology detection based on CBCT and panoramic...

Area of Science:

• Dentistry and Radiology
• Artificial Intelligence in Medical Imaging

Background:

• Panoramic radiography (PX) is crucial for dental diagnostics.
• Accurate tooth segmentation, numbering, and abnormality detection are vital for effective PX interpretation.
• Current methods may face challenges in precision and efficiency.

Purpose of the Study:

• To introduce PXseg, a novel approach for automated tooth segmentation, numbering, and abnormal morphology detection in PX.
• To enhance the performance of PXseg through optimized annotation and pre-training strategies.
• To evaluate the diagnostic efficiency of PXseg in clinical settings.

Main Methods:

• PXseg was developed using a combination of cone beam computed tomography (CBCT) data with 3D labels for pre-training and conventional PX (cPX) data with 2D labels for training.
• Internal datasets were used for visual and statistical analysis to assess segmentation and numbering.
• External datasets of cPXs with complex dental diseases evaluated abnormal tooth detection accuracy.
• Diagnostic testing compared efficiency with and without PXseg assistance.

Main Results:

• PXseg achieved high performance in tooth segmentation (DSC 0.882, F1-score 0.902) and numbering (F1-score 0.943), outperforming models without pre-training.
• Abnormal tooth morphology detection accuracy exceeded 0.957, a 4.3% improvement.
• A dedicated website was developed to aid PX interpretation, significantly enhancing diagnostic efficiency.

Conclusions:

• Pre-training with accurate labels from ctPX significantly improved PXseg's performance in segmentation, numbering, and abnormality detection.
• PXseg offers rapid and accurate results, streamlining the panoramic X-ray diagnostic workflow.
• PXseg demonstrates significant potential for clinical application in dentistry.