A comparison of GPT-4V's capability in optical coherence tomography images of age-related macular degeneration with expert assessments.

AIM: To examine the potential of GPT-4V(ision) in evaluating age-related macular degeneration (AMD) and other ophthalmological pathologies through optical coherence tomography (OCT) images.

METHODS

The study analyzed 90 OCT images, evenly divided among dry AMD, wet AMD, and normal control groups, using the GPT-4V model via the Julius AI platform. Each image was evaluated twice by the model and independently by two ophthalmologists (with a third arbitrating discrepancies) using a standardized set of 20 closed-ended questions covering critical parameters. The outputs were then compared statistically in terms of agreement, sensitivity, specificity, and predictive values to assess the model’s performance in detecting relevant pathological features.

RESULTS

The GPT-4V model exhibited agreement rates above 90%, high sensitivity, and specificity values in evaluating image quality, and normal OCT identification, detection of subretinal and intraretinal fluid, ellipsoid zone integrity, and wet-type AMD findings in AMD degeneration OCT images. Additionally, the two ophthalmology specialists exhibited excellent agreement (100%, p < 0.001) in 14 different questions when evaluating OCT images. However, the model’s performance was lower for dry AMD and normal OCT images, with sensitivities of 68.7% and 75.3%, respectively.

CONCLUSION

GPT-4V exhibited a promising performance as a clinical support tool in the evaluation of AMD through OCT images, albeit with limitations in detecting geographic atrophy and assessing dry AMD. However, the generalizabiity of these findings is constrained by the restricted dataset size and limited diagnostic spectrum. The GPT-4V model cannot therefore replace expert evaluation, although it appears promising as a clinical support tool. This study may also serve as a guide for larger-scale research.