Improving Laryngoscopy Image Analysis Through Integration of Global Information and Local Features in VoFoCD Dataset

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Otolaryngology

Background:

• Laryngoscopy is crucial for diagnosing vocal fold disorders, requiring precise identification of anatomical structures and lesions.
• Current diagnostic methods lack simultaneous optimization for object detection and image classification in laryngoscopy.
• Accurate visual assessment is vital for effective clinical decision-making in laryngology.

Purpose of the Study:

• To introduce the VoFoCD dataset for object detection and image classification in laryngoscopy.
• To propose a novel Multitask Efficient trAnsformer network for Laryngoscopy (MEAL) for enhanced diagnosis.
• To improve the accuracy and interpretability of AI-assisted vocal fold disorder diagnosis.

Main Methods:

• Development of the VoFoCD dataset comprising 1724 laryngology images across four classes and six glottic object types.
• Implementation of the MEAL network, a multitask transformer model for classifying vocal fold images and detecting glottic landmarks/lesions.
• Integration of attention maps within MEAL for explainable AI, visualizing critical regions for clinical interpretation.

Main Results:

• The MEAL model achieved high performance on the VoFoCD dataset, with 0.951 accuracy for image classification.
• Object detection performance was strong, yielding a mean average precision of 0.874 (mAP50).
• The model demonstrated robustness in simulated clinical scenarios, including those with laryngoscopy process variations.

Conclusions:

• The proposed MEAL network effectively integrates global and local features for accurate vocal fold image analysis.
• MEAL aids in visually identifying and classifying benign and malignant vocal fold lesions, supporting laryngologists.
• This AI-driven approach enhances diagnostic capabilities in laryngeal endoscopy, improving patient care for vocal fold disorders.