Comparing flow-based and anatomy-based features in the data-driven study of nasal pathologies

Andrea Schillaci, Kazuto Hasegawa, Carlotta Pipolo, Giacomo Boracchi, Maurizio Quadrio

Abstract

In several problems involving fluid flows, Computational Fluid Dynamics (CFD) provides detailed quantitative information, and often allows the designer to successfully optimize the system, by minimizing a cost function. Sometimes, however, one cannot improve the system with CFD alone, because a suitable cost function is not readily available: one notable example is diagnosis in medicine. The field of interest considered here is rhinology: a correct air flow is key for the functioning of the human nose, yet the notion of a functionally normal nose is not available, and a cost function cannot be written. An alternative and attractive pathway to diagnosis and surgery planning is offered by data-driven methods. In this work, we consider the machine-learning study of nasal pathologies caused by anatomic malformations, with the aim of understanding whether fluid dynamic features, available after a CFD analysis, are more effective than purely geometric features in the training of a neural network for regression. Our experiments are carried out on an extremely simplified anatomic model and a correspondingly simple CFD approach; nevertheless, they demonstrate that flow-based features perform better than geometry-based ones, and allow the training of a neural network with fewer inputs, a crucial advantage in fields like medicine.

Introduction

In rhinology, diagnosing and treating nasal pathologies like septal deviations remains a significant challenge, with surgical failure rates sometimes exceeding 50%. This difficulty stems from the lack of a clear definition of a "functionally normal" nose and the high anatomical variability among individuals. While surgeons traditionally rely on geometric assessments from CT scans, the nose's primary function is airflow. This paper investigates whether features derived from Computational Fluid Dynamics (CFD) simulations—which capture the non-linear relationship between anatomy and function—provide a more effective basis for machine learning-based diagnosis than purely geometric data.

Methodology

The study employs a synthetic parametric modeling approach to generate a controlled dataset of 200 nasal geometries. These include 100 "healthy" variations and 100 "pathological" cases mimicking turbinate hypertrophy. CFD simulations using OpenFOAM (RANS with $k-\omega$ SST) were performed to obtain flow fields for each geometry. To enable direct comparison across varying shapes, the researchers utilized "Functional Maps" to project surface data onto a common reference geometry.

Anatomical landmarks and cross-sectional flow patterns.

Feature extraction was performed using the Laplace–Beltrami operator to decompose surface data into eigenfunctions. The Lasso method then compressed this high-dimensional data into the 20 most informative coefficients for training Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN).

Dataset Generation

The synthetic dataset allows for precise labeling of pathology severity, which is often impossible with clinical data. The parametric model enables systematic variation of anatomical features, such as the position and size of the turbinates. This controlled environment is crucial for benchmarking the effectiveness of different feature types.

Simplified parametric CAD model with pathology variables.

Key Insights: The "Non-linear Filter" Effect

The core finding of this research is the "non-linear filter" effect of fluid dynamics. The Navier-Stokes equations act as a natural amplifier: subtle anatomical changes that might appear as noise in a geometric model produce distinct, large-scale signatures in flow-based fields like wall shear stress (WSS). Consequently, flow-based features (pressure and WSS) achieved prediction errors an order of magnitude lower than geometric features.

Results and Discussion

Wall Shear Stress (F2) emerged as the most potent predictor, with a test error of 0.019 mm in predicting swelling severity, compared to approximately 0.15 mm for geometric descriptors. This suggests that the physics of airflow inherently encodes the most relevant diagnostic information, allowing for more accurate models with fewer input parameters. Flow features were also found to be much better at ignoring "healthy" anatomical noise to focus on the actual pathology.

Conclusion

This work demonstrates that "physics-informed" features significantly enhance the performance of data-driven diagnostic tools in rhinology. By leveraging CFD to pre-process anatomical data into functional features, neural networks can more effectively distinguish between healthy variability and pathological malformation. Future work will focus on applying this methodology to real patient-specific geometries derived from clinical imaging, where the challenge will be labeling data without the benefit of parametric controls.

Summary updated

Abstract

In several problems involving fluid flows, Computational Fluid Dynamics (CFD) provides detailed quantitative information, and often allows the designer to successfully optimize the system, by minimizing a cost function. Sometimes, however, one cannot improve the system with CFD alone, because a suitable cost function is not readily available: one notable example is diagnosis in medicine. The field of interest considered here is rhinology: a correct air flow is key for the functioning of the human nose, yet the notion of a functionally normal nose is not available, and a cost function cannot be written. An alternative and attractive pathway to diagnosis and surgery planning is offered by data-driven methods. In this work, we consider the machine-learning study of nasal pathologies caused by anatomic malformations, with the aim of understanding whether fluid dynamic features, available after a CFD analysis, are more effective than purely geometric features in the training of a neural network for regression. Our experiments are carried out on an extremely simplified anatomic model and a correspondingly simple CFD approach; nevertheless, they demonstrate that flow-based features perform better than geometry-based ones, and allow the training of a neural network with fewer inputs, a crucial advantage in fields like medicine.
BibTeX
@inproceedings{schillaci2024comparing,
  author = {Andrea Schillaci and Kazuto Hasegawa and Carlotta Pipolo and Giacomo Boracchi and Maurizio Quadrio},
  title = {{Comparing flow-based and anatomy-based features in the data-driven study of nasal pathologies}},
  booktitle = {Flow},
  year = {2024},
  doi = {10.1017/flo.2024.3},
}
Figure