Generative-based Airway and Vessel Morphology Quantification on Chest CT Images

  • 2020-02-13 18:45:31
  • Pietro Nardelli, James C. Ross, Raúl San José Estépar
  • 7


Accurately and precisely characterizing the morphology of small pulmonarystructures from Computed Tomography (CT) images, such as airways and vessels,is becoming of great importance for diagnosis of pulmonary diseases. Thesmaller conducting airways are the major site of increased airflow resistancein chronic obstructive pulmonary disease (COPD), while accurately sizingvessels can help identify arterial and venous changes in lung regions that maydetermine future disorders. However, traditional methods are often limited dueto image resolution and artifacts. We propose a Convolutional Neural Regressor (CNR) that providescross-sectional measurement of airway lumen, airway wall thickness, and vesselradius. CNR is trained with data created by a generative model of syntheticstructures which is used in combination with Simulated and UnsupervisedGenerative Adversarial Network (SimGAN) to create simulated and refined airwaysand vessels with known ground-truth. For validation, we first use synthetically generated airways and vesselsproduced by the proposed generative model to compute the relative error anddirectly evaluate the accuracy of CNR in comparison with traditional methods.Then, in-vivo validation is performed by analyzing the association between thepercentage of the predicted forced expiratory volume in one second (FEV1\%) andthe value of the Pi10 parameter, two well-known measures of lung function andairway disease, for airways. For vessels, we assess the correlation between ourestimate of the small-vessel blood volume and the lungs' diffusing capacity forcarbon monoxide (DLCO). The results demonstrate that Convolutional Neural Networks (CNNs) provide apromising direction for accurately measuring vessels and airways on chest CTimages with physiological correlates.


Quick Read (beta)

loading the full paper ...