Machine learning assisted multiscale lung modeling to predict alveolar septal wall stress

Introduction:: Pathophysiologic conditions of the respiratory system, such as acute respiratory distress syndrome (ARDS) or ventilator-induced lung injury (VILI) can result in regional stiffening of the lung parenchyma and can alter the local biomechanical behavior. Such regional alterations are heterogeneous and patient-specific. Despite advances in quantifying temporal and spatial motion of the lungs using various dynamic imaging modalities (e.g., 4D computed tomography (4DCT)), the biomechanical cues triggering acute lung injury remain elusive. We hypothesize that there exists a threshold of alveolar septal wall stress (SWS) beyond which lung injury occurs. The calculation of SWS requires bridging the gap between lung function measurements at the macroscale and alveolar biomechanics at the microscale. However, multiscale lung models are computationally prohibitive, limiting the potential for clinical applications. To circumvent this impediment, we propose to use machine learning (ML) to assist with the prediction of SWS informed by kinematic information and pressure measurements.


Materials and Methods:: 4DCT scans were obtained for a healthy rat (n=1). Image registration was used to track the displacement and deformation of lung tissue. The process to quantify full 4D lung motion has been described in detail in our previous work [1]. To obtain the alveolar architecture, rat lungs were inflated to a constant airway pressure, before X-ray microtomography (µCT) imaging. A 1 mm x 1mm x 1mm cube was isolated, referred to as a representative tissue element (RTE). The RTE was reconstructed and used to perform in-silico experiments. The material in the cube was modeled as a neo-Hookean incompressible material. For the in-silico experiments, the RTE was stretched along three orthogonal directions (normal to the faces) where stretch values were chosen from Latin hypercube sampling bound by the volumetric deformation of J=0.7 and J=3 measured in lung motion quantification from 4DCT [1]. Three hundred simulations were performed by varying the stretch values along the three axes. A multi-layer perceptron (MLP) network was trained using the in-silico simulations with stretches (measured from 4DCT) and principal SWS extrema being input and output variables, respectively.

Results, Conclusions, and Discussions:: Results and Discussion: The stretches measured from 4DCT were converted to principal stretches and fed to the MLP network to obtain the alveolar-level stress parameters. Preliminary training results indicated that the accuracy of the training was greater than 88%. Based on the preliminary results, at a 15% stretch, mean stress of 1.5Pa, but maximum stress of 240Pa was observed in the RTE. This concentration of stress was consistent with prior studies [2] and highlights the need to include micromechanical behavior to accurately predict SWS. In future studies, the in-silico modeling will be extended to incorporate the effect of pulmonary surfactant and multiple RTEs with different alveolar architecture to train the ML model. The resulting ML model will be used to investigate the relationship between in-vivo SWS and lung injury.

Conclusions: In this study, we have presented an ML-assisted method to estimate alveolar septal wall stresses based on 4DCT imaging. We expect that SWS offers a sensitive metric to predict regions at high risk for lung injury during mechanical ventilation. In addition, we expect the ML model to be computationally efficient to be used in clinical settings.


Acknowledgements (Optional): : RA was supported by the NIH, under grant No. R00HL138288. DWK was supported in part by DoD PRMRP W81XWH-21-1-0507.

References (Optional): : References: [1] Neelakantan et al., Proc. SPIE 12466, MI 2023 [2] Concha, et al. (2018). J Mech Phys Solids 112: 126-144.