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 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 application considered here belongs to the field of 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 impairment 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 at 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 show 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.

The Challenge of Nasal Diagnosis

In rhinology, surgeons often decide on corrective procedures for nasal obstructions based on anatomical geometry observed in CT scans. However, the failure rate for these surgeries can exceed 50%. The fundamental challenge is that there is no universal definition of a "functionally normal" nose; anatomical variability is vast, and the relationship between the shape of the nasal cavity and the actual airflow is highly non-linear.

While CFD can simulate airflow, it does not automatically provide a diagnosis because we lack a clear "cost function" to minimize. Data-driven machine learning (ML) models offer a solution, but they typically require massive amounts of labeled data—which is scarce in medicine. This research asks: Can we train better ML models with less data by using CFD-derived flow features instead of raw geometry?

Methodology: From Geometry to Flow

The researchers developed a workflow that combines parametric modeling, fluid simulation, and geometric deep learning to overcome the lack of labeled medical data.

Comparison of realistic nasal anatomy (a) and the simplified parametric model (b) used for CFD simulations.
  1. Parametric Modeling: The team created a simplified, 3D parametric model of the human nasal cavity, generating 200 synthetic anatomies: 100 "healthy" and 100 "pathological" (simulating turbinate hypertrophy, a common swelling that blocks airflow).
  2. CFD Simulations: Steady-state inspiration simulations were performed using OpenFOAM (RANS equations with a $k-\omega$ SST turbulence model).
  3. Functional Maps (FM): To compare disparate shapes, the authors used Functional Maps to map all flow and geometric data back to a single "reference nose" using Laplace–Beltrami eigenfunctions. This allowed disparate shapes to be treated as a consistent dataset.
  4. Feature Extraction: They compared four types of input features: pointwise displacement and functional map sub-matrices (Geometry), versus wall pressure and wall shear stress magnitude (Flow).

The "Navier-Stokes Filter"

A profound insight from this study is that the Navier-Stokes equations act as a natural feature extractor. A tiny, localized geometric change—which might appear as "noise" to an ML model looking at raw geometry—can cause a large, systemic change in the flow field. This physics-based "filter" amplifies the relevant signals of the pathology. Furthermore, flow features tend to appear in the lower-order modes of the Laplace–Beltrami basis, meaning essential information is captured in fewer variables when looking at flow rather than geometry.

Results: Flow Features Outperform Geometry

The quantitative results were striking, showing that flow-based features consistently outperformed geometry-based features:

  • Superior Accuracy: For the most critical pathology parameters, the test error for flow features was roughly 10 times lower than for geometric features.
  • Best Feature: Wall shear stress (F2) proved to be the most informative feature, achieving the lowest average error across all tested pathologies (e.g., an error of 0.019 mm for pathology parameter $q_1$ vs. 0.156 mm for geometric distance).
  • Data Efficiency: The flow-based models achieved high correlation between predicted and true values even with a relatively small training set, suggesting that CFD data can significantly reduce the "data hunger" of medical ML models.

Conclusion and Future Work

By restricting flow features to the "wall" (the mucosal lining), the study remains biologically grounded, as the sensation of nasal patency is triggered by nerve endings in the lining. While the study used idealized CAD models, it demonstrates that flow-based features provide a more robust foundation for data-driven diagnosis. Future research will aim to apply this approach to real patient CT scans, moving toward more flexible, mesh-based analysis of actual clinical data.

Summary updated

Abstract

In several problems involving fluid flows, computational fluid dynamics (CFD) provides detailed quantitative information and 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 application considered here belongs to the field of 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 impairment 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 at 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 show 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