Background Oncocytomas of the ocular caruncle are rare benign epithelial tumors. Their clinical diagnosis is challenging, as they can mimic other benign or malignant lesions such as papilloma, nevus, squamous cell carcinoma, melanoma, or oncocytic carcinoma. For this reason, histopathological confirmation remains indispensable. The aim of this study was to test the ability of a multimodal large language model (ChatGPT, GPT-5, 2025 version) to generate diagnostic hypotheses directly from slit-lamp images, supported by brief clinical summaries. Case presentation We retrospectively analyzed two cases of caruncular oncocytoma that had undergone surgical excision with subsequent histopathological confirmation. For each case, ChatGPT was provided only with slit-lamp photographs of the lesion and a concise clinical summary including age, sex, and the site of the lesion (caruncle). No histopathological data or additional clinical details were supplied. In both cases, ChatGPT proposed oncocytoma as the primary diagnostic hypothesis. The model also generated differential diagnoses including papilloma, nevus, as well as the possibility of a malignant lesion such as squamous cell carcinoma or melanoma. Conclusions This proof-of-concept demonstrates, for the first time to our knowledge, that a general- purpose multimodal AI system can correctly recognize a rare ocular surface tumor from slit-lamp images. While preliminary and limited by the very small sample size, these findings suggest that large language models may assist clinicians in considering rare adnexal tumors during differential diagnosis. Further research on larger datasets is required, and histopathology will remain the gold standard for definitive diagnosis.
Sacchi et al. (Fri,) studied this question.