A Computational Geometry approach for Machine Learning based diagnosis of nasal breathing difficulties aided by CFD

Documento completo qui

Riccardo Margheritti

The work presents the pipeline that has been developed during the recent years, aimed at diagnosing nasal breathing difficulties through a Machine Learning (ML) model. Specifically, the ML model is designed to exploit CFD information extracted from simulations of patient-specific anatomies with pathologies. In dealing with shapes, a Computational Geometry (CG) approach has been adopted to automate some complicated procedures, and to perform Data Augmentation (DA) of the available dataset while ensuring consistent and well defined pathologies with unique labels.

The work is structured in three parts. In the first, starting from the complete CT scan of a healthy patient, a simplified geometry of the nasal cavities is extracted thanks to a CG tool, known as functional maps. The ability of the functional maps to create functional correspondences helps in performing DA, by increasing the number of available anatomies for which LES simulations are carried out. The second part describes the computational procedure for the LES simulations; the third part describes how CFD results can be compacted into a handful of values (called features) and used as input on a (pre-trained) Neural Network (NN), which performs inference on the pathologies. The pipeline has been previously trained on a database built with DA from 7 patients. The present work has provided data based on a 8th patient, used to test for the first time the accuracy of the classifier. Although the number of observations is still rather limited, the achieved accuracy of 80% is already satisfactory, and demonstrates the viability of the proposed approach.