Abstract View
Aid Pulmonary Disease Diagnosis and Treatment with CFD Modeling and Deep Learning: a New Perspective and Pilot Study
Changjie Cai, YU FENG, The University of Oklahoma Health Sciences Center
Abstract Number: 26
Working Group: Health-Related Aerosols
Abstract
According to the National Vital Statistic Report, chronic obstructive pulmonary disease (COPD) is the 4th leading cause of death in America, which causes severe breathing difficulty due to airway stiffening, loss of airway deformation capability, and airway blockage induced by inflammation. As the standard COPD treatment, inhalation of therapeutic nano-/micro- particles has illustrated a long-standing drug delivery barrier to achieving desired therapeutic outcomes, i.e., only approximately 25% of the drug particles can be delivered to the deeper lung with most of the particles depositing in the upper airway. To overcome such a barrier and increase the drug delivery efficiency from today’s 25% to 90% to the deeper lung for the better therapeutic outcome and reduced side effects, it is imperative to detect the obstruction locations. To achieve better treatment outcomes, it is also beneficial to detect the obstruction sites at earlier stages with noninvasive method. To pave the way to a noninvasive and automatic diagnostic method based on clinically measured intrathoracic medical images, we proposed and test the feasibility of a new noninvasive diagnostic methodology using both computational fluid dynamics (CFD) and binary Convolutional Neural Networks (CNN), i.e., “the expiratory airflow pattern analysis,” to identify the obstruction location in left or right lung deeper than generation 6 (G6) by automatic detection of the clinically measurable intrathoracic airflow velocity contour shifts using Hyperpolarized Magnetic Resonance Imaging (MRI). The virtual tracheobronchial (TB) tree employed in this study contains 44 small airway openings in total. To build the training data and test database, 1 of the 44 openings was blocked for each simulation, in order to mimic the least changes in obstruction conditions in human lung compared with the obstructions of multiple openings in left or right lung. We trained the YOLOv4, a state-of-the-art CNN mode to develop the prototype of the diagnosis algorithm, using 34 randomly selected images as the training inputs. The Mean Average Precision of the trained model for the other 10 images was 100% at a threshold (or probability of detection) of 0.5 with the Intersection over Union of 96.65%. However, in this study, only one subject-specific TB tree was employed in the CFD simulation, which neglects the inter-subject variability. Obstructions were assumed to only appears in either the left lung or the right lung in the training and test images, which did not contain the cases with obstructions in both lungs. To enhance the generalized predictability of the deep learning model in the near future, more subject-specific airway configurations will be reconstructed and employed in the CFD simulations to prepare the training and testing images with the effect of inter-subject variabilities.