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

Area of Science:

• Otolaryngology
• Artificial Intelligence in Medicine
• Medical Imaging Analysis

Background:

• Cholesteatoma detection relies heavily on radiologic interpretation.
• Large Language Models (LLMs) are emerging as potential tools for medical image analysis.
• Evaluating LLM performance in otologic diagnostics is crucial for clinical integration.

Purpose of the Study:

• To assess the diagnostic accuracy of the LLM Gemini 2.5 for cholesteatoma detection.
• To compare the LLM's performance against routine radiologic assessment.
• To explore the LLM's utility as a decision-support tool in otologic practice.

Main Methods:

• Retrospective analysis of 244 temporal bone MRIs (2017-2025).
• MRI data converted to video files for LLM evaluation without fine-tuning.
• Diagnostic metrics calculated against histopathology (gold standard).

Main Results:

• LLM achieved 79.1% diagnostic accuracy; radiology achieved 84.0%.
• Sensitivity was comparable, but specificity was higher for radiology.
• Performance remained consistent across disease subtypes and postoperative anatomy.

Conclusions:

• LLM demonstrated comparable diagnostic performance to radiology for cholesteatoma.
• LLMs offer rapid, image-centered analysis without specialized infrastructure.
• LLM-based systems can serve as practical complementary tools in otologic evaluation.