Signal Processing and Speech Communication Laboratory
hometheses & projects › Applying Semi-Supervised Learning to Adventitious Lung Sound Classification

Applying Semi-Supervised Learning to Adventitious Lung Sound Classification

Bachelor Project
Announcement date
01 Jan 2022
Florian Berger, Sebastian Pinter
Research Areas

** Abstract **

Analysis of respiratory sounds is an important factor for decisions regarding a patients disease and treatment. Respiratory sound classification utilizes machine learning and can help support such decisions by automation. In 2017, the International Conference on Biomedical and Health Informatics (ICBHI) challenge was released to the public. It provides a compiled audio database and a rule set for model comparison. Many competitive results for this challenge emerged from image-based neural networks, which use the transformation of audio data to the time-frequency domain.

In this thesis, we propose the application of Transformer-based neural networks in the field of adventitious lung sound classification for the first time. We combine different pre-processing methods from image-based networks and compare the impact on model performance. Following the official ICBHI rule set, we produce results by using Wav2Vec 2.0 amongst others. Furthermore, we compare these results with state-of-the-art models and conclude that our work performs slightly above average. We achieve harmonic scores of up to 50% and average scores of up to 54% by using the official ICBHI metrics.