Taming vision transformers for clinical laryngoscopy assessment

Area of Science:

• Otolaryngology
• Artificial Intelligence
• Medical Imaging

Background:

• Laryngoscopy is crucial for diagnosing laryngeal cancer (LCA), but suffers from high inter-observer variability and diagnostic challenges in differentiating precancerous from early cancerous lesions.
• Current diagnostic methods rely heavily on endoscopist expertise, leading to potential inconsistencies and delayed detection.

Purpose of the Study:

• To enhance laryngoscopic image analysis for improved early screening and detection of laryngeal cancer and precancerous conditions.
• To develop a robust AI model that overcomes data scarcity and achieves strong generalization capabilities.

Main Methods:

• Proposing MedFormer, a novel laryngeal cancer classification method utilizing the Vision Transformer (ViT) architecture.
• Implementing a customized transfer learning approach with pre-trained transformers to address data limitations and enable robust out-of-domain generalization.
• Fine-tuning a minimal set of additional parameters for efficient model adaptation.

Main Results:

• MedFormer achieved high sensitivity-specificity values: 98%-89% for precancerous lesions and 89%-97% for cancer detection.
• The model significantly outperformed Convolutional Neural Network (CNN) counterparts and other Vision Transformer (ViT)-based models.
• MedFormer matched or exceeded physician visual evaluations (PVE) in specific scenarios and demonstrated interpretability through visualizations.

Conclusions:

• Visual transformers show significant potential for clinical laryngoscopic assessments.
• MedFormer represents an effective AI-driven method for the early detection of laryngeal cancer, improving diagnostic accuracy and consistency.