Laryngeal Cancer Screening During Flexible Video Laryngoscopy Using Large Computer Vision Models.

Area of Science:

• Otolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• Laryngeal cancer diagnosis relies on flexible laryngoscopy.
• Objective interpretation of laryngoscopy findings can be challenging.
• AI offers potential to enhance diagnostic accuracy and efficiency.

Purpose of the Study:

• To develop an AI-assisted computer vision model for laryngeal cancer screening.
• To create a system for real-time annotation of flexible laryngoscopy videos.
• To classify lesions as potentially malignant or probably benign with confidence levels.

Main Methods:

• Developed computer vision models using laryngeal images and videos.
• Trained models to classify frames for usability and cancer screening.
• Implemented a separate model for lesion segmentation.
• Constructed a video stream annotation system integrating these models.

Main Results:

• The cancer screening model achieved 92% overall accuracy, 97.7% sensitivity, and 76.9% specificity.
• The lesion segmentation model attained an average precision of 0.595 (IoU 0.50).
• The model provides confidence levels for positive screening results.

Conclusions:

• The AI model demonstrates high sensitivity and adequate specificity for laryngeal cancer screening.
• Lesion segmentation aids endoscopists in identifying and describing potential abnormalities.
• Further optimization is needed for clinical deployment in real-time settings.