Rev Bras Oftalmol.2025;84:e0089
Assessment of diagnostic accuracy of ChatGPT in optical coherence tomography reports
DOI: 10.37039/1982.8551.20250089
ABSTRACT
Objective:
To assess the diagnostic accuracy of ChatGPT in analyzing OCT images.
Methods:
A total of 300 OCT images, previously interpreted by specialists, were analyzed by ChatGPT under three scenarios: image-only input; input with additional epidemiological data; and input with detailed clinical data. Sensitivity, specificity, and accuracy rates were compared across scenarios using appropriate statistical analyses (p < 0.05).
Results:
Overall accuracy was 66.33% in the image-only scenario, 82.33% with epidemiological data, and 92.67% with clinical data. Sensitivity for macular edema increased from 83.9% (scenario 1) to 100% (scenario 3). Specificity was 100% across all scenarios. Alterations such as epiretinal membrane and macular hole showed lower accuracy rates, even with additional data (91.5% and 95.4%, respectively).
Conclusion:
The results highlight the positive impact of contextual data integration on ChatGPT’s performance, corroborating previous studies on AI in medical diagnosis. Although the model is promising, it has limitations in analyzing more complex alterations and relies heavily on the quality of the data provided. Further studies are needed to validate its large-scale clinical application.
