Analysis of Upper Airway Morphology Using Four-Dimensional Dynamic MRI With Active Deep Learning-Based Automatic...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Respiratory System Anatomy

Background:

• Cine 4D MRI captures dynamic upper-airway morphology during breathing.
• Active-learning nnU-Net enhances segmentation accuracy while minimizing manual annotation efforts.

Purpose of the Study:

• To develop an automated method for upper airway segmentation using active learning on free-breathing cine 4D MRI.
• To quantify dynamic changes in upper airway morphology under different mouth positions.

Main Methods:

• Prospective cross-sectional study involving 84 adults (33 with symptoms).
• Utilized 3T, free-breathing TWIST sequence with closed- and open-mouth positions.
• Employed an active-learning nnU-Net model trained on radiologist-verified manual annotations.

Main Results:

• Achieved high segmentation accuracy (Dice 0.959 ± 0.019).
• Open-mouth breathing significantly altered airway length and reduced retropalatal cross-sectional area (CSA).
• Males and symptomatic individuals exhibited distinct airway volumes and CSA characteristics, with greater dynamic variability in symptomatic subjects.

Conclusions:

• Four-dimensional cine MRI combined with active-learning nnU-Net provides automated quantification of dynamic upper airway morphology.
• The study identified mouth position, symptoms, and sex as independent predictors of airway morphology and dynamics.