LarynxFormer: a transformer-based framework for processing and segmenting laryngeal images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Respiratory Medicine

Background:

• Manual diagnosis of exercise-induced laryngeal obstruction (EILO) is subjective and prone to human bias.
• Machine learning offers potential for objective, automated diagnosis of EILO through laryngeal image segmentation.
• Existing segmentation methods require comparison and improvement for clinical application.

Purpose of the Study:

• To develop and evaluate a novel machine learning framework, LarynxFormer, for objective EILO diagnosis.
• To compare the performance of transformer-based and convolutional-based models for laryngeal image segmentation.
• To assess the diagnostic accuracy and computational efficiency of automated laryngeal segmentation.

Main Methods:

• Implemented and trained four state-of-the-art segmentation models (convolutional and transformer-based) on a dataset of laryngeal images from continuous laryngoscopy exercise-tests (CLE-tests).
• Developed a new framework, LarynxFormer, incorporating pre-processing, transformer-based segmentation, and post-processing.
• Compared model performance using key metrics and computational speed.

Main Results:

• The proposed LarynxFormer framework demonstrated superior performance in laryngeal image segmentation.
• Transformer-based segmentation significantly outperformed conventional methods in accuracy and efficiency.
• Inference time was up to 2x faster with the transformer-based approach compared to other methods.

Conclusions:

• Machine learning, particularly transformer-based approaches, shows significant promise as an objective diagnostic tool for EILO.
• LarynxFormer provides an effective and efficient method for automated laryngeal segmentation, advancing EILO diagnostics.
• This study highlights the potential of AI to overcome limitations of manual EILO assessment.