AI-Powered Laryngoscopy: Exploring the Future with Google Gemini

Introduction: This study explores Google Gemini's ability to analyze laryngoscopy images and videos without prior training. It aims to determine if Gemini can identify and differentiate normal from pathological vocal folds, detect abnormalities, localize affected areas, and suggest diagnoses. By examining diverse laryngoscopy visuals, the study assesses the potential of foundational models in clinical decision support.

Methods: For patient privacy, images and videos of normal and pathologic vocal cords (VCs) were collected from Google Images and YouTube. YouTube videos were downloaded and edited with iMovie to remove identifying contextual information. Each image and video were individually uploaded to Google Gemini and Google AI Studio for analysis and disease classification. Various prompts were employed to get the AI to analyze the content, classify as normal or pathologic VCs, identify cord(s) affected, and specify the pathology present.

Results: Gemini demonstrated moderate accuracy in identifying VC pathology, including the location of lesions, particularly in high-resolution images with unobstructed views of VCs. Performance was best with normal VCs or polyps, while complex pathologies or lower-quality images posed challenges. Gemini displayed a large degree of inconsistency, occasionally providing different diagnoses during subsequent analyses of the same images or videos, despite using the same prompts.

Conclusions: Gemini demonstrates great potential for differentiating normal VCs from pathological ones, even classifying specific pathologies. While promising, in the current version, Gemini 1.5, accuracy varies with image quality and pathology complexity. It's occasional hesitancy to provide a diagnosis underscores the need for human expertise, especially in complex situations.

Limitations include inconsistent responses, reduced accuracy with lower-quality and more complex images and videos, and inability to analyze all videos. Future directions include comparing Gemini image and video analysis to other foundational models and assessing accuracy compared to untrained human novices.