Artificial intelligence-assisted multiscale lung modeling to predict alveolar septal wall stress

Injuries to the lung parenchyma, such as radiation-induced lung injury (RILI), can lead to heterogeneous ventilation and reduced lung function. The biomechanical drivers behind the onset and progression of such lung injuries are poorly understood. In this study, we developed a method based on tissue mechanical modeling and machine learning to estimate in-vivo alveolar septal wall stress, to understand the effect of parenchymal biomechanics on the onset and progression of lung injuries. Representative tissue elements (RTEs) of the lung parenchyma were reconstructed from X-ray microtomography imaging. A generative adversarial network (GAN) model was used to create synthetic RTEs to enhance the training data for an artificial neural network (ANN). Feature analysis indicated that the GAN model generated synthetic RTEs with features similar to those of the segmented RTEs. Finite element simulations were performed on both the segmented and synthetic RTEs and used to train the ANN. Stretch and geometric features served as inputs, and RTE stresses were the outputs. The ANN's testing accuracy was 61.6% before and 84.0% after including the synthetic RTEs. We used the ANN to investigate the evolution of stress in a rodent model of RILI. Reduced stress was observed in the lung tissue at 3-month post-radiation, indicating pneumonitis, followed by elevated stress at 5-month post-radiation, suggesting regional fibrosis. Further, stress heterogeneity at the 5-month timepoint indicated the presence of volutrauma. Overall, these results suggest that regional biomechanical markers can be used for early diagnosis and assessment of subclinical lung injuries that existing global measures cannot detect. Statement of Significance Injuries to the lung parenchymal tissue, such as radiation-induced lung injury (RILI), can lead to heterogeneous ventilation and reduced lung function. The drivers behind the onset and progression of such lung injuries are poorly understood. In this study, we developed a method to estimate alveolar septal wall stress (SWS) in vivo, through a combination of tissue mechanics, multiscale modeling, and machine learning (ML), to understand the effect of parenchymal biomechanics on the onset and progression of lung injuries. We applied the method to investigate a rodent model of RILI. The ML-based method could capture the pneumonitis and fibrosis behavior associated with early- and late-stage parenchymal remodeling post radiation, respectively. We expect that regional stresses in the lung parenchyma will provide key insights, including those obfuscated from global measures, into the onset and progression of lung injuries.