Documento completo qui
Kazuto Hasegawa
Computational fluid dynamics is expected to play a crucial role to diagnose a condition of human nose. The detailed information and solutions provided by numerical simulations enable us to not only analyze flow characteristics but also visualize flow fields in an understandable manner. However, it is also true that we often have to include expert knowledge to achieve a precise assessment since there are considerable uncertainties caused by the difficulty of diagnosis for human nose.
We here consider the use of data-driven frameworks for diagnosing nasal pathologies. Geometries and flow fields of 200 different noses, half of them exhibit some degree of turbinate hypertrophy, are utilized to predict their pathological parameters. First, the geometrical characteristics of the noses are extracted as a functional map between a reference healthy nose and other noses. The functional maps are then used to train neural networks and to predict pathological parameters. Three different machine learning models with different inputs and configurations are trained to examine the predictability of the methods for the pathologies. In addition to the prediction based on the geometric information, we also consider the utilization of flow measurements such as pressure and wall shear stress. The flow fields are considered as a function defined on nose which can be expressed by linear combinations of eigenfunctions of Laplace-Beltrami operator. The present results show reasonable agreements with the reference pathologies. We also find that the prediction performance can significantly be improved by including flow information as the input of machine learning models.