Multi-Instance Learning for Vocal Fold Leukoplakia Diagnosis Using White Light and Narrow-Band Imaging: A Multicenter...

Area of Science:

• Otolaryngology
• Artificial Intelligence in Medicine
• Medical Imaging

Background:

• Vocal fold leukoplakia (VFL) is a precancerous lesion that presents diagnostic challenges in laryngeal cancer detection.
• Accurate endoscopic diagnosis of VFL is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) model for distinguishing benign from malignant vocal fold leukoplakia (VFL).
• To assess the AI model's performance using both white light imaging (WLI) and narrow-band imaging (NBI) modalities.

Main Methods:

• A deep learning model combined with multi-instance learning (MIL) was trained on 7057 images from 426 patients.
• The model underwent internal and external validation using 1617 images from two additional hospitals.
• Real-time performance was evaluated on 50 videos, and a human-machine comparison involved 100 patients and 12 laryngologists.

Main Results:

• The AI model achieved high diagnostic performance with area under the receiver operating characteristic curve (AUC) values of 0.868 (internal) and 0.884 (external) validation.
• Video validation yielded an AUC of 0.825, demonstrating effectiveness in real-time scenarios.
• AI assistance significantly improved laryngologists' diagnostic AUC and accuracy, enhancing consistency.

Conclusions:

• A multicenter study successfully developed an AI model integrating WLI and NBI for VFL diagnosis.
• The AI model demonstrated effectiveness, particularly in assisting junior laryngologists.
• Further clinical validation and optimization are recommended to fully realize the AI's potential.