Cluster-Guided Multiscale Lung Modeling via Machine Learning

Accurate prediction of airflow distribution and aerosol transport in the human lungs, which are difficult to be measured in vivo but important to understand the structure and function relationship, is challenging. It is because the interplay between them spans more than two orders of magnitude in dimension from the trachea to alveoli. This chapter reviews the techniques and strategies for modeling lungs both within and between subjects, viz., subject specificity versus generalization from individuals to populations, with both exhibiting multiscale characteristics. For “within-subject” modeling, a computed tomography (CT)-derived subject-specific computational fluid dynamics (CFD) lung model is presented. The pipeline for building such an imaging-based lung model is composed of image segmentation and processing, geometrical modeling labeled with anatomical information, image registration, three-dimensional (accurate) and one-dimensional (approximate) coupling techniques, and a high-fidelity turbulent flow model. The subject-specific model is essential in predicting local structural and functional interactions. For “between subjects” modeling, machine learning is employed to identify homogeneous subpopulations (clusters), among healthy and diseased populations, aiming to bridge individual and population scales. For this purpose, three major issues need to be addressed. They are intersubject variability (due to, e.g., gender, age, and height), inter-site variability (due to scanner and imaging protocol differences), and definition of quantitative CT imaging-based metrics at multiple scales (due to alterations at different disease stages) needed for clustering. The use of the cluster membership to select representative subjects for detailed CFD analysis enables an examination of the cluster-specific structural and functional relationships.