Estimation of lung volume changes from frontal and lateral views of dynamic chest radiography using a convolutional neural network model: a computational phantom study

We aimed to estimate respiratory changes in lung volumes (Alung volume) using frontal and lateral dynamic chest radiography (DCR) by employing a convolutional neural network (CNN) learning approach trained and tested using the four-dimensional (4D) extended cardiac-torso (XCAT) phantom. Twenty XCAT phantoms of males (5 normal, 5 overweight, and 5 obese) and females (5 normal) were generated to obtain 4D computed tomography (CT) of a virtual patient. XCAT phantoms were projected in frontal and lateral directions. We estimated lung volumes of the XCAT phantoms using CNN learning techniques. One dataset consisted of a right- or left-half frontal view, a lateral view, and ground truth (GT) knowledge of each phantom in the same respiratory phase. Alung volume were calculated by subtracting the lung volume estimated at the maximum exhale from that at the maximal inhale, and was compared with Alung volume calculated from the known GT. Alung volume was successfully estimated from frontal and lateral DCR images of XCAT phantoms by a CNN learning approach. There was a correlation for Alung volume between GT and estimation in both lungs. There were no significant differences in the estimation error between the right and left lungs, males and females, and males having different physiques. We confirmed that DCR has potential use in the estimation of Alung volume, which corresponds to vital capacity (VC) in pulmonary function tests (PFT). Pulmonary function could be assessed by DCR even in patients with infectious diseases who can't do PFT using a spirometer.