Exercise-induced Laryngeal Obstruction Diagnostics Using Machine Learning

Abstract

Exercise-induced laryngeal obstruction (EILO), characterized by laryngeal narrowing during physical exercise, poses a significant challenge, especially for athletes and active youth, impacting performance and quality of life. The continuous laryngoscopy exercise test (CLE-test) is currently the gold standard for assessing EILO. This procedure involves filming the larynx with a laryngoscope. The test is often followed by an evaluation by a clinician, scoring the patient’s severity of EILO. This score corresponds to the amount of adduction of the laryngeal structures.

Manual scoring methods for EILO diagnostics preserve challenges, mainly the problem of subjectivity. Several studies have proposed machine learning (ML) methods for image segmentation of the laryngeal structures, aiming to develop more efficient and objective diagnostic tools. In recent years, ML, particularly deep learning, has made significant advancements and presents a promising area for further exploration in EILO diagnostics.

This thesis explores state-of-the-art ML approaches for segmenting laryngeal structures and proposes a new framework to improve segmentation performance and efficiency. Two scientific papers have been developed: Paper A explores the current manual methods for assessing EILO and cutting-edge ML methods focusing on segmenting the laryngeal structures. Paper B compares these ML methods with transformer-based approaches. Our findings indicate that transformer-based segmentation outperforms the efficiency of the current methods. Additionally, the paper introduces a framework named LarynxFormer, which comprises pre-processing, segmentation, and post-processing of laryngeal images.