CFD-driven Machine Learning for the Classification of Nasal Breathing Difficulties - Thesis

Aug 2, 2026

CFD-driven Machine Learning provides a promising support approach for assessing nasal breathing difficulties by enabling automatic classification of nasal pathologies. The method builds a training dataset by running fluid-dynamic simulations on virtual surgeries performed on healthy anatomical noses extracted from CT scans, then refines the pipeline to reduce computational cost while improving accuracy. Comparative evaluation contrasts reduced-order model data against Large Eddy Simulations for binary and four-class tasks. New three-class classification is enabled by the expanded database, and feature selection further improves scores, making reduced-order inference suitable for near real-time analysis.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages86

Reading Time3 h 37 min

AccessFree to download

Introduction

Nasal Anatomy, Pathologies, and 3D Reconstruction from CT Scans

Anatomy and Physiology of the Nasal Cavity

Gathering Geometries from CT Scans

Pathology Insertion

Database Construction

Theoretical Background

Shape

CFD-driven Machine Learning for the Classification of Nasal Breathing Difficulties Tesi di Laurea Magistrale in Aeronautical Engineering-Ingegneria Aeronautica Author: Carlantonio di Venosa Student ID: 221973 Advisor: Prof. Maurizio Quadrio Co-advisors: Niccolò Berizzi, Angelo Raimondo Favero Academic Year: 2024-25 “Ma Nino non aver paura Di sbagliare un calcio di rigore Non è mica da questi particolari Che si giudica un giocatore Un giocatore lo vedi dal coraggio Dall’altruismo e dalla fantasia” —Francesco De Gregori, La leva calcistica della classe ’68’ i Abstract CFD-driven Machine Learning (ML) represents a powerful and promising approach to assist in medical assessment of nasal breathing difficulties. In the context, previous works have proposed a procedure for an automatic classification of nasal pathologies by leveraging CFD-data in a ML framework. For training the classifier, fluid-dynamic data are gathered by running simulations on a database of pathological noses. The database has been constructed by virtual surgeries on healthy anatomical noses extracted from CT scans. This Thesis study presents a revised version of the pipeline, with modifications introduced to enhance computational efficiency and result accuracy. In particular, the procedure for the construction of the database has been modified, reducing the computational cost. Moreover, the new procedure has been leveraged for enriching the database with new pathological noses (from 270 to 784) . From a methodological and operational perspective, a comparative evaluation was conducted: a classifier trained on data generated by a reduced-order model was compared with a classifier trained on data obtained from Large Eddy Simulations (LES) . The analysis covers two tasks: a binary classification (distinguishing between septal deviationsand turbinate hypertrophies) and a four-class classification (discriminating among anterior/posterior septal deviation and inferior/middle turbinate hypertrophy) . In terms of accuracy, the reduced-order model data outscored (0.98) the LES-data (0.85) for the binary classification. The multi-class problem results, instead, showed similar performance. This is a key finding, since the computational cost and the time requested for a simulation with the reduced-order model are incredibly low (few seconds), thus making it particularly suitable for a real-time type of analysis. A new type of classification has been carried out by leveraging the new database, distinguishing among three classes − septal deviation, turbinate hypertrophy and both combined. Eventually, it was shown that the adoption of a feature selection technique has the po- ii | Abstract tential to greatly improve the scores of the classification (e.g. accuracy from 0.65 to 0.86 for the four-class problem) . Keywords: nasal breathing difficulties, machine learning, computational fluid-dynamics iii Abstract in lingua italiana Il Machine Learning (ML) guidato dalla fluidodinamica computazionale (CFD) rappresenta un promettente strumento di supporto per le valutazioni cliniche delle difficoltà respiratorie nasali. In questo contesto, lavori precedenti hanno utilizzato una procedura per la classificazione automatica delle patologie nasali sfruttando i dati prodotti dalla CFD in ambito ML. Correntemente, per allenare il classificatore, vengono raccolti i datida simulazioni su un database di nasi patologici costruito inserendo virtualmente le deformazioni corrispondenti alle malattie su nasi sani, estratti via TAC. Il presente lavoro di Tesi rivisita questa procedura con l’obiettivo di ridurre il costo computazionale e migliorare l’accuratezza nei risultati. In particolare, il metodo per la creazione del database è stato reso più efficiente da un punto di vista computazionale; lo stesso metodo è stato, inoltre, utilizzato per la creazionedi nuovi nasi patologici (da 270 a 784) . Dal punto di vista metodologico-operativo, si è ope

FAQs

How is the training data for nasal classification generated?

Fluid-dynamic data are obtained by running simulations on a database built from virtual surgeries applied to healthy anatomical noses extracted from CT scans.

What models are compared in the classifier evaluation?

A classifier trained on reduced-order model data is compared against a classifier trained on data obtained from Large Eddy Simulations (LES), across binary and four-class tasks.

How do the reduced-order model and feature selection affect performance and practicality?

The reduced-order model achieves higher binary accuracy than LES while requiring only a few seconds per simulation, and feature selection can substantially improve classification accuracy for the multi-class problem.

Aug 2, 2026

Upload by Theodora