Deep learning to predict respiratory lung-tissue displacement from multi-phase pulmonary computed tomography images

Abstract

Lung surgery, particularly lung puncture, is challenging because respiration motion significantly displaces lung tissues (e.g., blood vessels, trachea), including targeted lesions. To improve the success of lung surgery, it is essential to determine how lung tissues move during respiration. Early work used simple mathematical models (e.g., linear models) to predict lung-tissue displacement. However, the quality and quantity of pulmonary computed tomography (CT) images used to build such models were limited, preventing their universal application. Moreover, such mathematical models are too simple to describe how lung tissues are displaced during respiration. In this work, we analyzed the relationship between body-surface and lung-tissue displacement in a large number of multi-phase pulmonary CT images to create a model predicting lung-tissue displacement. First, a systematic method based on unsupervised learning approaches was designed to calculate the displacement of the body surface and lung tissues. Second, a deep neural network model was developed to predict lung-tissue displacement, using the body-surface displacement observed in CT imaging data as input. Multi-phase pulmonary CT images from 199 patients were collected to train and test our model. The validation experiment demonstrated that the prediction accuracy based on our deep learning model reached 79.8%, which outperformed traditional machine learning approaches like the linear model (79.2%), linear support vector regression model (79.2%), and support vector regression model with radial basis function (67.2%). These results demonstrated that our method could predict lung-tissue displacement based on body-surface displacement during respiration, potentially improving the success of lung puncture procedures.