Evaluation of ChatGPT's ability to detect intraosseous pathologies with panoramic radiographs.

BACKGROUND

To evaluate the diagnostic performance of GPT-4 V and GPT-5 in detecting intraosseous jaw lesions on panoramic radiographs and to assess the readability of their responses.

METHODS

119 biopsy-confirmed panoramic radiographs were retrospectively analyzed. Each image was presented to both models with four standardized diagnostic questions. Unguided and region-guided prompting were compared. Information accuracy, relevance, actionability, quality, and readability (Flesch Reading Ease Score and Flesch–Kincaid Grade Level) were recorded.

RESULTS

Unguided lesion detection sensitivity was low. With location guidance, lesion identification increased to 73.9% for GPT-4 V and 78.1% for GPT-5; however, diagnostic sensitivity remained limited (31.1% and 36.1%). Information accuracy, relevance, actionability, and quality scores were mostly concentrated in the lower-to-middle range (scores 2–3). No statistically significant differences were found between the models (p > 0.05). Readability analysis showed high textual complexity (mean FKGL 12.8 and 12.7).

CONCLUSION

GPT-4 V and GPT-5 did not achieve clinically acceptable diagnostic sensitivity on panoramic radiographs. Further model refinement is required before diagnostic use.