Deep Learning-Based Generation of 4D Reference Lung Volumes for Personalized Respiratory Function Assessment

Lung function exhibits substantial variability both between and within individuals, necessitating a comprehensive understanding of both static and dynamic aspects to enable more personalized and precise respiratory disease assessments. Effective diagnosis, treatment, and follow-up critically depend on reliable reference points. This study aims to establish a personalized reference lung dynamic, serving as a baseline for individual patients. By leveraging dynamic 3D magnetic resonance lung images, we aim to predict a patient's "healthy" hypothetical respiratory dynamics, enabling a comparison of their actual function and dynamics with their own predicted baseline. To this effect, we employ an AI-based technique to generate dynamic lung models. Our approach combines two neural networks: a convolutional autoencoder that extracts latent vectors from lung volumes, allowing us to generate a static lung representation based on real data, and an LSTM-based recurrent neural network that captures temporal dependencies in respiratory deformations to generate a healthy respiratory motion.