Upper Airway Classification in Sleep Endoscopy Examinations using Convolutional Recurrent Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Sleep Medicine

Background:

• Drug-induced sleep endoscopy (DISE) is crucial for diagnosing obstructive sleep apnea (OSA) by identifying upper airway (UA) collapse.
• The VOTE classification system categorizes UA collapse at the velum (V), oropharynx (O), tongue (T), and epiglottis (E).
• Analyzing DISE videos is challenging due to anatomical variations and video distortions.

Purpose of the Study:

• To develop and validate an automated method for classifying upper airway regions in DISE videos.
• To establish a foundational step towards AI-driven analysis of DISE for OSA diagnosis.

Main Methods:

• A convolutional recurrent neural network (CRNN) with ResNet18 and bidirectional LSTM was employed.
• The model was trained and tested on 24 annotated DISE videos, classifying 5-second video sequences.
• A three-class problem was defined: velum (V), oropharynx/tongue/epiglottis (OTE), and indeterminate (X).

Main Results:

• The CRNN model achieved an overall accuracy of 82% and an F1-score of 79% for the three-class classification.
• The model demonstrated potential in recognizing UA regions across diverse patient anatomies.
• These findings support the feasibility of using AI for automated DISE video analysis.

Conclusions:

• Automated classification of UA regions in DISE videos is achievable with current AI technology.
• This approach shows promise for future development of AI tools to predict collapse severity and type.
• Ultimately, this could lead to automated diagnoses derived directly from DISE videos, improving OSA management.