Large Language Models for Cholesteatoma Diagnosis: A Pathology-Validated Study

To evaluate the diagnostic performance of the large language model (LLM) Gemini 2.5 for cholesteatoma detection using histopathology as the reference standard, and to compare its performance with that of routine radiologic assessment. A secondary objective was to assess the model's consistency across clinically relevant subgroups and its potential utility as a complementary, image-centered decision-support tool for otologic practice.

In this retrospective diagnostic accuracy study, 244 patients who underwent temporal bone MRI, including non-echo-planar diffusion-weighted imaging, between 2017 and 2025 were analyzed. MRI data sets were converted into multislice video files and evaluated by an LLM without fine-tuning or task-specific training. Diagnostic performance metrics were calculated using histopathology as the gold standard and compared using the McNemar test. Subgroup analyses were performed for primary versus recurrent cholesteatoma and according to mastoidectomy history.

Histopathology confirmed cholesteatoma in 121 patients. Overall diagnostic accuracy was 79.1% for the LLM and 84.0% for radiology. Sensitivity did not differ significantly between the 2 approaches, whereas specificity was higher for radiologic assessment. Diagnostic performance remained stable across primary and recurrent disease and in postoperative anatomy. The model also identified an incidental intracranial cystic lesion outside the temporal bone.

The large language model (LLM) demonstrated diagnostic performance comparable to routine radiologic interpretation and provided rapid, image-centered analysis without the need for training or specialized infrastructure. These findings suggest that LLM-based systems may serve as practical complementary tools in cholesteatoma evaluation, particularly as an adjunct to radiologic assessment and preoperative decision-making in otologic practice.

Level III.