Artificial Intelligence Versus Human Dental Expertise in Diagnosing Periapical Pathosis on Periapical Radiographs: A...

Area of Science:

• Dental Radiology
• Artificial Intelligence in Medicine
• Diagnostic Accuracy

Background:

• Periapical pathosis diagnosis is crucial for endodontic success but challenged by 2D imaging limitations and interpretation subjectivity.
• Artificial intelligence (AI) presents a potential solution, yet its diagnostic granularity compared to human clinicians requires thorough investigation.
• Evaluating AI's role in interpreting periapical radiographs is essential for advancing diagnostic capabilities.

Purpose of the Study:

• To assess the diagnostic accuracy of ChatGPT-5 in identifying periapical radiographic abnormalities.
• To compare the performance of ChatGPT-5 against a consensus of three board-certified oral radiologists.
• To investigate the agreement and predictors of correct AI classification in periapical radiograph interpretation.

Main Methods:

• A retrospective diagnostic accuracy study involving 270 periapical radiographs.
• Independent interpretation of radiographs by ChatGPT-5 using a standardized prompt and by a three-expert oral radiologist consensus.
• Comparison of diagnostic accuracy, agreement (Cohen's kappa), and classification predictors between AI and expert consensus.

Main Results:

• ChatGPT-5 achieved high sensitivity (87.5%) but low specificity (12.5%), yielding 50.0% overall diagnostic accuracy.
• The AI model exhibited a tendency to over-identify pathology, classifying 87.5% of radiographs as abnormal versus 50.0% by experts.
• Agreement was near-perfect for anatomical localization (κ = 0.857) but poor for abnormality detection (κ = 0.000); significant disagreement noted for lesion border characterization (κ = 0.127).

Conclusions:

• ChatGPT-5 demonstrates high sensitivity in visual interpretation of periapical radiographs but limited specificity and inconsistent morphological feature characterization hinder its clinical diagnostic reliability.
• The AI system systematically over-diagnoses and describes lesions differently than dental experts, categorizing them as more structurally defined.
• AI holds potential as a sensitive initial screening tool, but findings necessitate validation by dental professionals to mitigate false positives and ensure accurate characterization.