Multicenter Clinical Validation of an Artificial Intelligence Diagnostic Classification Model for Laryngoscopy Images

Area of Science:

• Otolaryngology
• Artificial Intelligence
• Medical Imaging

Background:

• Laryngeal lesion classification is crucial for patient management.
• Accurate differentiation between high-risk (HR) and low-risk (LR) lesions is essential.
• Existing diagnostic methods can be subjective and resource-intensive.

Purpose of the Study:

• To develop and externally validate an AI-powered computer-aided diagnosis (CADx) model.
• To classify laryngeal lesions from laryngoscopy images as high-risk (HR) or low-risk (LR).
• To assess the generalizability and diagnostic performance of the CADx model across diverse populations.

Main Methods:

• Retrospective multicenter development and validation of a CADx model using over 20,000 laryngoscopy images.
• Deep learning model trained on internal data and evaluated on independent Greek and Spanish cohorts.
• Performance measured by accuracy, AUC, and comparison with otolaryngologists, general practitioners, and ChatGPT-4o.

Main Results:

• The CADx model achieved high performance across internal (accuracy/AUC: 0.90/0.89) and external datasets (Greek: 0.85/0.85; Spanish: 0.88/0.88).
• Model accuracy was statistically noninferior to otolaryngologists and expert laryngologists.
• AI model outperformed general practitioners and ChatGPT-4o in classifying laryngeal lesions.

Conclusions:

• A large-scale multicenter validation confirms the generalizability of the AI CADx model for laryngeal endoscopy.
• The model demonstrates performance comparable to clinicians in discriminating between low-risk and high-risk lesions.
• The AI tool shows potential to enhance diagnostic capabilities, particularly in resource-limited settings.