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

Manual diagnostic methods for assessing exercise-induced laryngeal obstruction (EILO) contain human bias and can lead to subjective decisions. Several studies have proposed machine learning methods for segmenting laryngeal structures to automate and make diagnostic outcomes more objective. Four state-of-the-art models for laryngeal image segmentation are implemented, trained, and compared using our pre-processed dataset containing laryngeal images derived from continuous laryngoscopy exercise-test (CLE-test) data. These models include both convolutional-based and transformer-based methods. We propose a new framework called LarynxFormer, consisting of a pre-processing pipeline, transformer-based segmentation, and post-processing of laryngeal images. This study contributes to the investigation of using machine learning as a diagnostic tool for EILO. Furthermore, we show that a transformer-based approach for larynx segmentation outperforms conventional state-of-the-art image segmentation methods in terms of performance metrics and computational speed, demonstrating up to 2x faster inference time compared to the other methods.