Comparison between observer-based and AI-based reading of CBCT datasets: An interrater-reliability study

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• Cone beam computed tomography (CBCT) is crucial for diagnosing periodontal diseases.
• Convolutional neural networks (CNNs) show promise in medical image analysis.
• Evaluating AI performance against human experts is essential for clinical adoption.

Purpose of the Study:

• To compare the diagnostic accuracy of human observers and a CNN in identifying periodontal lesions on CBCT images.
• To assess the interobserver agreement among human readers for periodontal lesion detection.

Main Methods:

• 38 CBCT datasets were analyzed by three calibrated general dentists and a deep CNN (Diagnocat).
• Human readers identified vertical bone loss, furcation involvement, and periradicular osteolysis using 3D MPR mode.
• The CNN performed an automated analysis of the same datasets.

Main Results:

• Human observers achieved significantly better performance than the CNN in detecting periodontal lesions.
• Good to very good interobserver agreement was observed among human readers.
• Moderate agreement was noted for vertical lesion detection among dentists.

Conclusions:

• The CNN evaluated was ineffective for detecting periodontal lesions in CBCT datasets.
• The CNN requires further training to be a reliable tool for automated periodontal lesion evaluation.
• Human expertise remains superior for interpreting CBCT scans for periodontal pathologies.