Automatic detection of pathological breath sounds to assess bronchial congestion

Period, duration

january - july 2024

Profile required

Students in 4th or 5th year of a Master's degree or engineering school in IT

Gratuity

4.35€/h (standard)

Location

Annecy-le-vieux ; Maison de la mécatronique ; Symme laboratory

Subject

Background:
MUKROBS is a clinical study in which we record pathological respiratory sounds at different sites in the thoracic cavity.
During the clinical study, we are collecting acoustic data from different patients, for which we are seeking to identify and count 3 main types of sound (crackling, sibilant, purring). We are in the process of collecting thousands of raw recordings of different durations, from a pathological group and a control group, for a total of 60 subjects.
This internship follows a first two-month internship, which enabled us to lay the foundations for our work, in particular by using the ICBHI database to develop an initial AI.

Missions:
- Complete and exploit the state of the art on existing machine learning models and signal processing techniques for adventitious lung noise detection.
- If necessary, continue the implementation of a ResNet 50 network, with layer optimization.
- Establish traceability of tests and parameterizations to enable statistical analysis of performance.
- Transfer learning: The next step is to build our own database based on the clinical study measurements. In the end, we'll have to apply these model(s) to our own data.

Profile required:

- Autonomy
- Curiosity
- Dynamic
- Machine Learning
- Signal processing
- Matlab / Python
- PyTorch/TensorFlow
- Knowledge of wavelets or prosody-based sound and speech processing would be a plus.

Contact

laurent.goujon
@univ-smb.fr

Project team

Marine Loubet, SYMME doctoral student
Laurent Goujon, teacher-researcher IUT Annecy / SYMME
Argheesh Bhanot, teacher-researcher IUT Annecy / LISTIC
Christine Barthod, teacher-researcher IUT Annecy / SYMME