A single latent channel is sufficient for biomedical glottis segmentation

Area of Science:

• Medical Imaging
• Deep Learning
• Laryngology

Background:

• Glottis segmentation is vital for analyzing laryngeal high-speed videoendoscopy.
• Deep neural networks offer automatic glottis segmentation but lack transparency.
• Understanding network internals is crucial for clinical adoption.

Purpose of the Study:

• To investigate the minimal requirements for effective glottis segmentation using deep neural networks.
• To enhance the interpretability and efficiency of deep learning models for laryngeal imaging.

Main Methods:

• Systematic ablation studies were performed on deep segmentation networks.
• Analysis of the latent space properties and its relationship to glottal area.
• Evaluation of encoding and decoding strategies for the latent representation.

Main Results:

• A single latent channel bottleneck is sufficient for accurate glottal area segmentation.
• The latent space represents glottal segmentation through three pixel subtypes, enabling transparent interpretation.
• The latent space correlates with glottal area waveform, is efficiently encoded (4 bits), and decoded with high accuracy.

Conclusions:

• Glottis segmentation can be achieved with highly optimized, efficient, and explainable deep neural networks.
• These findings facilitate clinical acceptance and application of AI in laryngeal examinations.
• Online deep learning-assisted monitoring holds significant potential for future laryngeal diagnostics.