Enhancing Diagnostic Accuracy in Laryngology: Leveraging Real Laryngoscope Images for AI Model Training

Introduction: The diagnosis of laryngeal pathologies is challenging due to the subjectivity of traditional methods like flexible laryngoscopy. This study investigates whether incorporating AI-generated synthetic laryngoscope images into training datasets can improve the diagnostic accuracy and robustness of AI models in identifying laryngeal conditions, addressing data limitations in AI applications for laryngology.

Methods: This retrospective study utilized real laryngoscopic images from de-identified patient records at Albany Medical Center, including patients diagnosed with laryngeal cancer, polyps, nodules, cysts, Reinke's edema, papillomas, leukoplakia, and normal anatomy. ICD10 diagnostic codes yielded 50 laryngoscopic images from each pathology and were categorized into groups. A latent diffusion model generated synthetic images based on these datasets and were used to train an AI model capable of classifying laryngeal pathology. The performance of the diagnostic model was compared between two groups: one trained on a combination of real and synthetic images, and the other trained solely with real images, to determine whether incorporating synthetic images enhanced the model's accuracy. Image accuracy and model robustness were assessed via subjective analysis and the Fréchet Inception Distance (FID), which compares the distribution of AI-generated images with real images based on feature extraction.

Results: The AI model trained with both real and synthetic images shows improvements in diagnostic accuracy and robustness compared to the model trained solely with real images. FID analysis indicates that the synthetic images appear to closely match the distribution of real images, supporting their potential usefulness in model training.

Conclusions: Incorporating synthetic laryngoscopic images into AI-model training improves diagnostic accuracy and robustness, offering a promising solution to data limitations in AI applications in laryngology. Further validation in diverse clinical settings is required to assess the generalizability of synthetic image application in clinical practice.