GlottisNetV2: Temporal Glottal Midline Detection Using Deep Convolutional Neural Networks.

Area of Science:

• Otolaryngology
• Medical Imaging
• Artificial Intelligence

Background:

• High-speed videoendoscopy is vital for quantitative laryngology.
• Accurate glottis segmentation and glottal midline detection are essential for vocal fold parameter calculation.
• Existing automated methods for glottal midline detection lack clinical applicability.

Purpose of the Study:

• To develop an advanced multitask deep neural network for precise glottis segmentation and glottal midline detection.
• To improve the accuracy and reliability of automated glottal midline estimation in endoscopic images.
• To provide a clinically applicable tool for quantitative laryngology analysis.

Main Methods:

• Development of a dual-decoder deep neural network (GlottisNetV2) utilizing pose estimation techniques.
• Training and evaluation of the network using the BAGLS dataset and a custom temporal dataset.
• Implementation of hyperparameter tuning and temporal filtering for enhanced prediction accuracy.

Main Results:

• GlottisNetV2 demonstrated superior performance over GlottisNet, with a Mean Absolute Percentage Error (MAPE) ranging from 1.85% to 6.3%.
• The model achieved faster convergence and improved median prediction accuracy from 2.1% to 1.76% using 12 consecutive frames and temporal filtering.
• The proposed architecture enables stable and reliable glottal midline predictions suitable for clinical use.

Conclusions:

• The developed GlottisNetV2 offers accurate and stable glottal midline prediction, crucial for quantitative laryngology.
• This deep learning approach enhances the clinical utility of high-speed videoendoscopy for vocal fold analysis.
• The method facilitates reliable analysis of symmetry measures and supports diagnosis of laryngeal conditions.