Documento completo qui
This thesis is part of OpenNOSE, a multidisciplinary research project aiming at the development of an open source, patient-specific tool to support medical personnel in the diagnoses and surgical procedures to the nasal cavity, to reduce costs and collateral damage. The specific task concerns the exploration of a novel approach to the evaluation of nasal airflow quality, based on ML (Machine Learning) techniques.
The first step is the creation of a simplified CAD (Computer-Aided Design) model of the nasal cavity. Secondly, some meaningful geometrical parameters are defined, in close collaboration with the otolaryngology experts at San Paolo Hospital of Milan, and then combined to generate 200 different nasal geometries, half of which exhibit some version and degree of turbinate hypertrophy. RANS (Reynolds-Averaged Navier-Stokes) simulations allow the reproduction of the airflow inside the models during the inhalation phase. Some flow features are extracted from the raw CFD (Computational Fluid Dynamics) results, summarising the entire flow fields in a few numerical values to employ as the predictive model input. The quality of the chosen features is then tested in a feature selection process, based on ML techniques. Additionally, this procedure allows the gathering of insights about the non-trivial connection between model geometry and the airflow inside it. Lastly, a final regression model assesses the real predictive value of the chosen features.